{ "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com", "method": "https://localrigcheck.com/methodology/#lab", "note": "One record per llama-bench process; withdrawn and corrected records stay, with their corrections. The raw output each record names in its raw field (llama-bench JSON, host offer, logs) is kept in RigCheck's own repository and is available on request.", "records": [ { "schema": 1, "id": "lab-2026-10-21t131403z-16gb-m39330-gemma-4-26b-a4b-ud-iq3-xxs-32k-r1", "status": "current", "label": "RigCheck measurement", "identity": { "model": "gemma-4-26b-a4b", "repo": "unsloth/gemma-4-26B-A4B-it-GGUF", "revision": "c099eb48e663fd284577b04978a94ffccb261841", "file": "gemma-4-26B-A4B-it-UD-IQ3_XXS.gguf", "quant": "UD-IQ3_XXS", "bytes": 11416548832, "sha256": "3b7aec683e176e1d957d007e0fb590611d82aec4db358d99b7f9714b6bc7bfc2" }, "environment": { "provider": "Vast.ai", "machine_id": 39330, "host_id": 98690, "gpu": "NVIDIA GeForce RTX 4060 Ti", "gpu_memory_mib": 16380, "gpu_count": 1, "driver": "580.173.02", "cuda_driver_api": "13.0", "power_limit_w": 165.0, "power_limit_note": null, "pcie": { "gen_max": 3, "width_max": 8, "listed_bandwidth_gbps": 6.0 }, "cpu": "Intel(R) Core(TM) i5-8500 CPU @ 3.00GHz", "cpu_cores_allocated": 6.0, "ram_gib_allocated": 31.2, "host_cpu_threads": 6, "host_ram_gib": 31.2, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m gemma-4-26B-A4B-it-UD-IQ3_XXS.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 32256 -r 3 -o json", "context_target": 32768, "context_note": "Prompt processing: 512 tokens timed after 32256 untimed tokens (context 32768 requested). Generation: 128 tokens timed after the same 32256 untimed tokens (context 32384 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 6, "repetitions": 3, "run": 1, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 32256, "context_requested": 32768, "mean_tps": 1561.5, "stdev_tps": 26.25, "samples_tps": [ 1565.37, 1533.53, 1585.6 ] }, "generation": { "timed_tokens": 128, "start_depth": 32256, "context_requested": 32384, "mean_tps": 65.33, "stdev_tps": 0.14, "samples_tps": [ 65.17, 65.41, 65.43 ] }, "gpu_memory_peak_mib": 12572, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 10, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T13:47:12Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T131403Z-16gb-m39330/gemma-4-26b-a4b-32768-r1.json", "corrections": [], "page": "https://localrigcheck.com/gemma-4/" }, { "schema": 1, "id": "lab-2026-10-21t131403z-16gb-m39330-gemma-4-26b-a4b-ud-iq3-xxs-32k-r2", "status": "current", "label": "RigCheck measurement", "identity": { "model": "gemma-4-26b-a4b", "repo": "unsloth/gemma-4-26B-A4B-it-GGUF", "revision": "c099eb48e663fd284577b04978a94ffccb261841", "file": "gemma-4-26B-A4B-it-UD-IQ3_XXS.gguf", "quant": "UD-IQ3_XXS", "bytes": 11416548832, "sha256": "3b7aec683e176e1d957d007e0fb590611d82aec4db358d99b7f9714b6bc7bfc2" }, "environment": { "provider": "Vast.ai", "machine_id": 39330, "host_id": 98690, "gpu": "NVIDIA GeForce RTX 4060 Ti", "gpu_memory_mib": 16380, "gpu_count": 1, "driver": "580.173.02", "cuda_driver_api": "13.0", "power_limit_w": 165.0, "power_limit_note": null, "pcie": { "gen_max": 3, "width_max": 8, "listed_bandwidth_gbps": 6.0 }, "cpu": "Intel(R) Core(TM) i5-8500 CPU @ 3.00GHz", "cpu_cores_allocated": 6.0, "ram_gib_allocated": 31.2, "host_cpu_threads": 6, "host_ram_gib": 31.2, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m gemma-4-26B-A4B-it-UD-IQ3_XXS.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 32256 -r 3 -o json", "context_target": 32768, "context_note": "Prompt processing: 512 tokens timed after 32256 untimed tokens (context 32768 requested). Generation: 128 tokens timed after the same 32256 untimed tokens (context 32384 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 6, "repetitions": 3, "run": 2, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 32256, "context_requested": 32768, "mean_tps": 1557.8, "stdev_tps": 25.55, "samples_tps": [ 1565.77, 1529.21, 1578.4 ] }, "generation": { "timed_tokens": 128, "start_depth": 32256, "context_requested": 32384, "mean_tps": 65.33, "stdev_tps": 0.16, "samples_tps": [ 65.14, 65.42, 65.42 ] }, "gpu_memory_peak_mib": 12572, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 10, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T13:47:39Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T131403Z-16gb-m39330/gemma-4-26b-a4b-32768-r2.json", "corrections": [], "page": "https://localrigcheck.com/gemma-4/" }, { "schema": 1, "id": "lab-2026-10-21t131403z-16gb-m39330-gemma-4-26b-a4b-ud-iq3-xxs-32k-r3", "status": "current", "label": "RigCheck measurement", "identity": { "model": "gemma-4-26b-a4b", "repo": "unsloth/gemma-4-26B-A4B-it-GGUF", "revision": "c099eb48e663fd284577b04978a94ffccb261841", "file": "gemma-4-26B-A4B-it-UD-IQ3_XXS.gguf", "quant": "UD-IQ3_XXS", "bytes": 11416548832, "sha256": "3b7aec683e176e1d957d007e0fb590611d82aec4db358d99b7f9714b6bc7bfc2" }, "environment": { "provider": "Vast.ai", "machine_id": 39330, "host_id": 98690, "gpu": "NVIDIA GeForce RTX 4060 Ti", "gpu_memory_mib": 16380, "gpu_count": 1, "driver": "580.173.02", "cuda_driver_api": "13.0", "power_limit_w": 165.0, "power_limit_note": null, "pcie": { "gen_max": 3, "width_max": 8, "listed_bandwidth_gbps": 6.0 }, "cpu": "Intel(R) Core(TM) i5-8500 CPU @ 3.00GHz", "cpu_cores_allocated": 6.0, "ram_gib_allocated": 31.2, "host_cpu_threads": 6, "host_ram_gib": 31.2, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m gemma-4-26B-A4B-it-UD-IQ3_XXS.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 32256 -r 3 -o json", "context_target": 32768, "context_note": "Prompt processing: 512 tokens timed after 32256 untimed tokens (context 32768 requested). Generation: 128 tokens timed after the same 32256 untimed tokens (context 32384 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 6, "repetitions": 3, "run": 3, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 32256, "context_requested": 32768, "mean_tps": 1561.05, "stdev_tps": 25.54, "samples_tps": [ 1557.85, 1537.26, 1588.04 ] }, "generation": { "timed_tokens": 128, "start_depth": 32256, "context_requested": 32384, "mean_tps": 65.23, "stdev_tps": 0.13, "samples_tps": [ 65.09, 65.35, 65.25 ] }, "gpu_memory_peak_mib": 12572, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 10, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T13:48:06Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T131403Z-16gb-m39330/gemma-4-26b-a4b-32768-r3.json", "corrections": [], "page": "https://localrigcheck.com/gemma-4/" }, { "schema": 1, "id": "lab-2026-10-21t131403z-16gb-m39330-gemma-4-26b-a4b-ud-iq3-xxs-4k-r1", "status": "current", "label": "RigCheck measurement", "identity": { "model": "gemma-4-26b-a4b", "repo": "unsloth/gemma-4-26B-A4B-it-GGUF", "revision": "c099eb48e663fd284577b04978a94ffccb261841", "file": "gemma-4-26B-A4B-it-UD-IQ3_XXS.gguf", "quant": "UD-IQ3_XXS", "bytes": 11416548832, "sha256": "3b7aec683e176e1d957d007e0fb590611d82aec4db358d99b7f9714b6bc7bfc2" }, "environment": { "provider": "Vast.ai", "machine_id": 39330, "host_id": 98690, "gpu": "NVIDIA GeForce RTX 4060 Ti", "gpu_memory_mib": 16380, "gpu_count": 1, "driver": "580.173.02", "cuda_driver_api": "13.0", "power_limit_w": 165.0, "power_limit_note": null, "pcie": { "gen_max": 3, "width_max": 8, "listed_bandwidth_gbps": 6.0 }, "cpu": "Intel(R) Core(TM) i5-8500 CPU @ 3.00GHz", "cpu_cores_allocated": 6.0, "ram_gib_allocated": 31.2, "host_cpu_threads": 6, "host_ram_gib": 31.2, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m gemma-4-26B-A4B-it-UD-IQ3_XXS.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 3584 -r 3 -o json", "context_target": 4096, "context_note": "Prompt processing: 512 tokens timed after 3584 untimed tokens (context 4096 requested). Generation: 128 tokens timed after the same 3584 untimed tokens (context 3712 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 6, "repetitions": 3, "run": 1, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 3584, "context_requested": 4096, "mean_tps": 2907.07, "stdev_tps": 90.99, "samples_tps": [ 2945.38, 2803.19, 2972.64 ] }, "generation": { "timed_tokens": 128, "start_depth": 3584, "context_requested": 3712, "mean_tps": 76.51, "stdev_tps": 0.24, "samples_tps": [ 76.24, 76.66, 76.63 ] }, "gpu_memory_peak_mib": 11984, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 10, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T13:46:22Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T131403Z-16gb-m39330/gemma-4-26b-a4b-4096-r1.json", "corrections": [], "page": "https://localrigcheck.com/gemma-4/" }, { "schema": 1, "id": "lab-2026-10-21t131403z-16gb-m39330-gemma-4-26b-a4b-ud-iq3-xxs-4k-r2", "status": "current", "label": "RigCheck measurement", "identity": { "model": "gemma-4-26b-a4b", "repo": "unsloth/gemma-4-26B-A4B-it-GGUF", "revision": "c099eb48e663fd284577b04978a94ffccb261841", "file": "gemma-4-26B-A4B-it-UD-IQ3_XXS.gguf", "quant": "UD-IQ3_XXS", "bytes": 11416548832, "sha256": "3b7aec683e176e1d957d007e0fb590611d82aec4db358d99b7f9714b6bc7bfc2" }, "environment": { "provider": "Vast.ai", "machine_id": 39330, "host_id": 98690, "gpu": "NVIDIA GeForce RTX 4060 Ti", "gpu_memory_mib": 16380, "gpu_count": 1, "driver": "580.173.02", "cuda_driver_api": "13.0", "power_limit_w": 165.0, "power_limit_note": null, "pcie": { "gen_max": 3, "width_max": 8, "listed_bandwidth_gbps": 6.0 }, "cpu": "Intel(R) Core(TM) i5-8500 CPU @ 3.00GHz", "cpu_cores_allocated": 6.0, "ram_gib_allocated": 31.2, "host_cpu_threads": 6, "host_ram_gib": 31.2, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m gemma-4-26B-A4B-it-UD-IQ3_XXS.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 3584 -r 3 -o json", "context_target": 4096, "context_note": "Prompt processing: 512 tokens timed after 3584 untimed tokens (context 4096 requested). Generation: 128 tokens timed after the same 3584 untimed tokens (context 3712 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 6, "repetitions": 3, "run": 2, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 3584, "context_requested": 4096, "mean_tps": 2908.07, "stdev_tps": 99.33, "samples_tps": [ 2937.74, 2797.29, 2989.18 ] }, "generation": { "timed_tokens": 128, "start_depth": 3584, "context_requested": 3712, "mean_tps": 76.48, "stdev_tps": 0.24, "samples_tps": [ 76.21, 76.62, 76.61 ] }, "gpu_memory_peak_mib": 11984, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 10, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T13:46:33Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T131403Z-16gb-m39330/gemma-4-26b-a4b-4096-r2.json", "corrections": [], "page": "https://localrigcheck.com/gemma-4/" }, { "schema": 1, "id": "lab-2026-10-21t131403z-16gb-m39330-gemma-4-26b-a4b-ud-iq3-xxs-4k-r3", "status": "current", "label": "RigCheck measurement", "identity": { "model": "gemma-4-26b-a4b", "repo": "unsloth/gemma-4-26B-A4B-it-GGUF", "revision": "c099eb48e663fd284577b04978a94ffccb261841", "file": "gemma-4-26B-A4B-it-UD-IQ3_XXS.gguf", "quant": "UD-IQ3_XXS", "bytes": 11416548832, "sha256": "3b7aec683e176e1d957d007e0fb590611d82aec4db358d99b7f9714b6bc7bfc2" }, "environment": { "provider": "Vast.ai", "machine_id": 39330, "host_id": 98690, "gpu": "NVIDIA GeForce RTX 4060 Ti", "gpu_memory_mib": 16380, "gpu_count": 1, "driver": "580.173.02", "cuda_driver_api": "13.0", "power_limit_w": 165.0, "power_limit_note": null, "pcie": { "gen_max": 3, "width_max": 8, "listed_bandwidth_gbps": 6.0 }, "cpu": "Intel(R) Core(TM) i5-8500 CPU @ 3.00GHz", "cpu_cores_allocated": 6.0, "ram_gib_allocated": 31.2, "host_cpu_threads": 6, "host_ram_gib": 31.2, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m gemma-4-26B-A4B-it-UD-IQ3_XXS.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 3584 -r 3 -o json", "context_target": 4096, "context_note": "Prompt processing: 512 tokens timed after 3584 untimed tokens (context 4096 requested). Generation: 128 tokens timed after the same 3584 untimed tokens (context 3712 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 6, "repetitions": 3, "run": 3, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 3584, "context_requested": 4096, "mean_tps": 2914.97, "stdev_tps": 102.16, "samples_tps": [ 2950.81, 2799.73, 2994.38 ] }, "generation": { "timed_tokens": 128, "start_depth": 3584, "context_requested": 3712, "mean_tps": 76.51, "stdev_tps": 0.25, "samples_tps": [ 76.22, 76.66, 76.66 ] }, "gpu_memory_peak_mib": 11984, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 10, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T13:46:45Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T131403Z-16gb-m39330/gemma-4-26b-a4b-4096-r3.json", "corrections": [], "page": "https://localrigcheck.com/gemma-4/" }, { "schema": 1, "id": "lab-2026-10-21t131403z-16gb-m39330-qwen-3-8-27b-ud-iq3-xxs-32k-r1", "status": "current", "label": "RigCheck measurement", "identity": { "model": "qwen-3-8-27b", "repo": "unsloth/Qwen3.8-27B-GGUF", "revision": "4ca720788d1e01f1bff70c033e0d0028fd02e502", "file": "Qwen3.8-27B-UD-IQ3_XXS.gguf", "quant": "UD-IQ3_XXS", "bytes": 10934860704, "sha256": "c0b7c3038681ed2e3040456c1dd45f9858b6c2290bed172c70388a94874f3eee" }, "environment": { "provider": "Vast.ai", "machine_id": 39330, "host_id": 98690, "gpu": "NVIDIA GeForce RTX 4060 Ti", "gpu_memory_mib": 16380, "gpu_count": 1, "driver": "580.173.02", "cuda_driver_api": "13.0", "power_limit_w": 165.0, "power_limit_note": null, "pcie": { "gen_max": 3, "width_max": 8, "listed_bandwidth_gbps": 6.0 }, "cpu": "Intel(R) Core(TM) i5-8500 CPU @ 3.00GHz", "cpu_cores_allocated": 6.0, "ram_gib_allocated": 31.2, "host_cpu_threads": 6, "host_ram_gib": 31.2, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m Qwen3.8-27B-UD-IQ3_XXS.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 32256 -r 3 -o json", "context_target": 32768, "context_note": "Prompt processing: 512 tokens timed after 32256 untimed tokens (context 32768 requested). Generation: 128 tokens timed after the same 32256 untimed tokens (context 32384 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 6, "repetitions": 3, "run": 1, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 32256, "context_requested": 32768, "mean_tps": 578.55, "stdev_tps": 5.3, "samples_tps": [ 580.48, 572.56, 582.62 ] }, "generation": { "timed_tokens": 128, "start_depth": 32256, "context_requested": 32384, "mean_tps": 19.77, "stdev_tps": 0.02, "samples_tps": [ 19.74, 19.78, 19.78 ] }, "gpu_memory_peak_mib": 12820, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 10, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T13:36:28Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T131403Z-16gb-m39330/qwen-3-8-27b-32768-r1.json", "corrections": [], "page": "https://localrigcheck.com/qwen-3-8-27b/" }, { "schema": 1, "id": "lab-2026-10-21t131403z-16gb-m39330-qwen-3-8-27b-ud-iq3-xxs-32k-r2", "status": "current", "label": "RigCheck measurement", "identity": { "model": "qwen-3-8-27b", "repo": "unsloth/Qwen3.8-27B-GGUF", "revision": "4ca720788d1e01f1bff70c033e0d0028fd02e502", "file": "Qwen3.8-27B-UD-IQ3_XXS.gguf", "quant": "UD-IQ3_XXS", "bytes": 10934860704, "sha256": "c0b7c3038681ed2e3040456c1dd45f9858b6c2290bed172c70388a94874f3eee" }, "environment": { "provider": "Vast.ai", "machine_id": 39330, "host_id": 98690, "gpu": "NVIDIA GeForce RTX 4060 Ti", "gpu_memory_mib": 16380, "gpu_count": 1, "driver": "580.173.02", "cuda_driver_api": "13.0", "power_limit_w": 165.0, "power_limit_note": null, "pcie": { "gen_max": 3, "width_max": 8, "listed_bandwidth_gbps": 6.0 }, "cpu": "Intel(R) Core(TM) i5-8500 CPU @ 3.00GHz", "cpu_cores_allocated": 6.0, "ram_gib_allocated": 31.2, "host_cpu_threads": 6, "host_ram_gib": 31.2, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m Qwen3.8-27B-UD-IQ3_XXS.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 32256 -r 3 -o json", "context_target": 32768, "context_note": "Prompt processing: 512 tokens timed after 32256 untimed tokens (context 32768 requested). Generation: 128 tokens timed after the same 32256 untimed tokens (context 32384 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 6, "repetitions": 3, "run": 2, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 32256, "context_requested": 32768, "mean_tps": 577.75, "stdev_tps": 5.01, "samples_tps": [ 579.08, 572.21, 581.96 ] }, "generation": { "timed_tokens": 128, "start_depth": 32256, "context_requested": 32384, "mean_tps": 19.77, "stdev_tps": 0.02, "samples_tps": [ 19.74, 19.78, 19.78 ] }, "gpu_memory_peak_mib": 12820, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 10, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T13:37:39Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T131403Z-16gb-m39330/qwen-3-8-27b-32768-r2.json", "corrections": [], "page": "https://localrigcheck.com/qwen-3-8-27b/" }, { "schema": 1, "id": "lab-2026-10-21t131403z-16gb-m39330-qwen-3-8-27b-ud-iq3-xxs-32k-r3", "status": "current", "label": "RigCheck measurement", "identity": { "model": "qwen-3-8-27b", "repo": "unsloth/Qwen3.8-27B-GGUF", "revision": "4ca720788d1e01f1bff70c033e0d0028fd02e502", "file": "Qwen3.8-27B-UD-IQ3_XXS.gguf", "quant": "UD-IQ3_XXS", "bytes": 10934860704, "sha256": "c0b7c3038681ed2e3040456c1dd45f9858b6c2290bed172c70388a94874f3eee" }, "environment": { "provider": "Vast.ai", "machine_id": 39330, "host_id": 98690, "gpu": "NVIDIA GeForce RTX 4060 Ti", "gpu_memory_mib": 16380, "gpu_count": 1, "driver": "580.173.02", "cuda_driver_api": "13.0", "power_limit_w": 165.0, "power_limit_note": null, "pcie": { "gen_max": 3, "width_max": 8, "listed_bandwidth_gbps": 6.0 }, "cpu": "Intel(R) Core(TM) i5-8500 CPU @ 3.00GHz", "cpu_cores_allocated": 6.0, "ram_gib_allocated": 31.2, "host_cpu_threads": 6, "host_ram_gib": 31.2, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m Qwen3.8-27B-UD-IQ3_XXS.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 32256 -r 3 -o json", "context_target": 32768, "context_note": "Prompt processing: 512 tokens timed after 32256 untimed tokens (context 32768 requested). Generation: 128 tokens timed after the same 32256 untimed tokens (context 32384 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 6, "repetitions": 3, "run": 3, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 32256, "context_requested": 32768, "mean_tps": 577.94, "stdev_tps": 5.1, "samples_tps": [ 580.63, 572.06, 581.12 ] }, "generation": { "timed_tokens": 128, "start_depth": 32256, "context_requested": 32384, "mean_tps": 19.77, "stdev_tps": 0.02, "samples_tps": [ 19.75, 19.78, 19.78 ] }, "gpu_memory_peak_mib": 12820, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 10, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T13:38:51Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T131403Z-16gb-m39330/qwen-3-8-27b-32768-r3.json", "corrections": [], "page": "https://localrigcheck.com/qwen-3-8-27b/" }, { "schema": 1, "id": "lab-2026-10-21t131403z-16gb-m39330-qwen-3-8-27b-ud-iq3-xxs-4k-r1", "status": "current", "label": "RigCheck measurement", "identity": { "model": "qwen-3-8-27b", "repo": "unsloth/Qwen3.8-27B-GGUF", "revision": "4ca720788d1e01f1bff70c033e0d0028fd02e502", "file": "Qwen3.8-27B-UD-IQ3_XXS.gguf", "quant": "UD-IQ3_XXS", "bytes": 10934860704, "sha256": "c0b7c3038681ed2e3040456c1dd45f9858b6c2290bed172c70388a94874f3eee" }, "environment": { "provider": "Vast.ai", "machine_id": 39330, "host_id": 98690, "gpu": "NVIDIA GeForce RTX 4060 Ti", "gpu_memory_mib": 16380, "gpu_count": 1, "driver": "580.173.02", "cuda_driver_api": "13.0", "power_limit_w": 165.0, "power_limit_note": null, "pcie": { "gen_max": 3, "width_max": 8, "listed_bandwidth_gbps": 6.0 }, "cpu": "Intel(R) Core(TM) i5-8500 CPU @ 3.00GHz", "cpu_cores_allocated": 6.0, "ram_gib_allocated": 31.2, "host_cpu_threads": 6, "host_ram_gib": 31.2, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m Qwen3.8-27B-UD-IQ3_XXS.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 3584 -r 3 -o json", "context_target": 4096, "context_note": "Prompt processing: 512 tokens timed after 3584 untimed tokens (context 4096 requested). Generation: 128 tokens timed after the same 3584 untimed tokens (context 3712 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 6, "repetitions": 3, "run": 1, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 3584, "context_requested": 4096, "mean_tps": 946.12, "stdev_tps": 15.23, "samples_tps": [ 953.38, 928.61, 956.36 ] }, "generation": { "timed_tokens": 128, "start_depth": 3584, "context_requested": 3712, "mean_tps": 22.92, "stdev_tps": 0.03, "samples_tps": [ 22.88, 22.94, 22.94 ] }, "gpu_memory_peak_mib": 11000, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 10, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T13:34:22Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T131403Z-16gb-m39330/qwen-3-8-27b-4096-r1.json", "corrections": [], "page": "https://localrigcheck.com/qwen-3-8-27b/" }, { "schema": 1, "id": "lab-2026-10-21t131403z-16gb-m39330-qwen-3-8-27b-ud-iq3-xxs-4k-r2", "status": "current", "label": "RigCheck measurement", "identity": { "model": "qwen-3-8-27b", "repo": "unsloth/Qwen3.8-27B-GGUF", "revision": "4ca720788d1e01f1bff70c033e0d0028fd02e502", "file": "Qwen3.8-27B-UD-IQ3_XXS.gguf", "quant": "UD-IQ3_XXS", "bytes": 10934860704, "sha256": "c0b7c3038681ed2e3040456c1dd45f9858b6c2290bed172c70388a94874f3eee" }, "environment": { "provider": "Vast.ai", "machine_id": 39330, "host_id": 98690, "gpu": "NVIDIA GeForce RTX 4060 Ti", "gpu_memory_mib": 16380, "gpu_count": 1, "driver": "580.173.02", "cuda_driver_api": "13.0", "power_limit_w": 165.0, "power_limit_note": null, "pcie": { "gen_max": 3, "width_max": 8, "listed_bandwidth_gbps": 6.0 }, "cpu": "Intel(R) Core(TM) i5-8500 CPU @ 3.00GHz", "cpu_cores_allocated": 6.0, "ram_gib_allocated": 31.2, "host_cpu_threads": 6, "host_ram_gib": 31.2, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m Qwen3.8-27B-UD-IQ3_XXS.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 3584 -r 3 -o json", "context_target": 4096, "context_note": "Prompt processing: 512 tokens timed after 3584 untimed tokens (context 4096 requested). Generation: 128 tokens timed after the same 3584 untimed tokens (context 3712 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 6, "repetitions": 3, "run": 2, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 3584, "context_requested": 4096, "mean_tps": 945.21, "stdev_tps": 15.3, "samples_tps": [ 952.22, 927.66, 955.74 ] }, "generation": { "timed_tokens": 128, "start_depth": 3584, "context_requested": 3712, "mean_tps": 22.92, "stdev_tps": 0.03, "samples_tps": [ 22.88, 22.94, 22.94 ] }, "gpu_memory_peak_mib": 11000, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 10, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T13:34:49Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T131403Z-16gb-m39330/qwen-3-8-27b-4096-r2.json", "corrections": [], "page": "https://localrigcheck.com/qwen-3-8-27b/" }, { "schema": 1, "id": "lab-2026-10-21t131403z-16gb-m39330-qwen-3-8-27b-ud-iq3-xxs-4k-r3", "status": "current", "label": "RigCheck measurement", "identity": { "model": "qwen-3-8-27b", "repo": "unsloth/Qwen3.8-27B-GGUF", "revision": "4ca720788d1e01f1bff70c033e0d0028fd02e502", "file": "Qwen3.8-27B-UD-IQ3_XXS.gguf", "quant": "UD-IQ3_XXS", "bytes": 10934860704, "sha256": "c0b7c3038681ed2e3040456c1dd45f9858b6c2290bed172c70388a94874f3eee" }, "environment": { "provider": "Vast.ai", "machine_id": 39330, "host_id": 98690, "gpu": "NVIDIA GeForce RTX 4060 Ti", "gpu_memory_mib": 16380, "gpu_count": 1, "driver": "580.173.02", "cuda_driver_api": "13.0", "power_limit_w": 165.0, "power_limit_note": null, "pcie": { "gen_max": 3, "width_max": 8, "listed_bandwidth_gbps": 6.0 }, "cpu": "Intel(R) Core(TM) i5-8500 CPU @ 3.00GHz", "cpu_cores_allocated": 6.0, "ram_gib_allocated": 31.2, "host_cpu_threads": 6, "host_ram_gib": 31.2, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m Qwen3.8-27B-UD-IQ3_XXS.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 3584 -r 3 -o json", "context_target": 4096, "context_note": "Prompt processing: 512 tokens timed after 3584 untimed tokens (context 4096 requested). Generation: 128 tokens timed after the same 3584 untimed tokens (context 3712 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 6, "repetitions": 3, "run": 3, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 3584, "context_requested": 4096, "mean_tps": 945.26, "stdev_tps": 14.72, "samples_tps": [ 950.74, 928.59, 956.46 ] }, "generation": { "timed_tokens": 128, "start_depth": 3584, "context_requested": 3712, "mean_tps": 22.92, "stdev_tps": 0.03, "samples_tps": [ 22.88, 22.94, 22.94 ] }, "gpu_memory_peak_mib": 11000, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 10, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T13:35:16Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T131403Z-16gb-m39330/qwen-3-8-27b-4096-r3.json", "corrections": [], "page": "https://localrigcheck.com/qwen-3-8-27b/" }, { "schema": 1, "id": "lab-2026-10-21t135058z-4090-m47598-gemma-4-26b-a4b-ud-q4-k-xl-32k-r1", "status": "current", "label": "RigCheck measurement", "identity": { "model": "gemma-4-26b-a4b", "repo": "unsloth/gemma-4-26B-A4B-it-GGUF", "revision": "c099eb48e663fd284577b04978a94ffccb261841", "file": "gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf", "quant": "UD-Q4_K_XL", "bytes": 17010980576, "sha256": "ef728c8e0c337fd1067b947af006e38a9ef2419e56feced4fd29b4bf0636e30c" }, "environment": { "provider": "Vast.ai", "machine_id": 47598, "host_id": 55116, "gpu": "NVIDIA GeForce RTX 4090", "gpu_memory_mib": 24564, "gpu_count": 1, "driver": "580.142", "cuda_driver_api": "13.0", "power_limit_w": 400.0, "power_limit_note": "host power limit 400 W; the card ships at 450 W", "pcie": { "gen_max": 4, "width_max": 16, "listed_bandwidth_gbps": 18.6 }, "cpu": "AMD EPYC 7402 24-Core Processor", "cpu_cores_allocated": 12.0, "ram_gib_allocated": 62.9, "host_cpu_threads": 48, "host_ram_gib": 251.5, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 32256 -r 3 -o json", "context_target": 32768, "context_note": "Prompt processing: 512 tokens timed after 32256 untimed tokens (context 32768 requested). Generation: 128 tokens timed after the same 32256 untimed tokens (context 32384 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 24, "repetitions": 3, "run": 1, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 32256, "context_requested": 32768, "mean_tps": 6232.58, "stdev_tps": 457.54, "samples_tps": [ 6304.33, 5743.4, 6650 ] }, "generation": { "timed_tokens": 128, "start_depth": 32256, "context_requested": 32384, "mean_tps": 146.31, "stdev_tps": 0.95, "samples_tps": [ 145.22, 146.89, 146.83 ] }, "gpu_memory_peak_mib": 18197, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 0, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T14:05:40Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T135058Z-4090-m47598/gemma-4-26b-a4b-32768-r1.json", "corrections": [], "page": "https://localrigcheck.com/gemma-4/" }, { "schema": 1, "id": "lab-2026-10-21t135058z-4090-m47598-gemma-4-26b-a4b-ud-q4-k-xl-32k-r2", "status": "current", "label": "RigCheck measurement", "identity": { "model": "gemma-4-26b-a4b", "repo": "unsloth/gemma-4-26B-A4B-it-GGUF", "revision": "c099eb48e663fd284577b04978a94ffccb261841", "file": "gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf", "quant": "UD-Q4_K_XL", "bytes": 17010980576, "sha256": "ef728c8e0c337fd1067b947af006e38a9ef2419e56feced4fd29b4bf0636e30c" }, "environment": { "provider": "Vast.ai", "machine_id": 47598, "host_id": 55116, "gpu": "NVIDIA GeForce RTX 4090", "gpu_memory_mib": 24564, "gpu_count": 1, "driver": "580.142", "cuda_driver_api": "13.0", "power_limit_w": 400.0, "power_limit_note": "host power limit 400 W; the card ships at 450 W", "pcie": { "gen_max": 4, "width_max": 16, "listed_bandwidth_gbps": 18.6 }, "cpu": "AMD EPYC 7402 24-Core Processor", "cpu_cores_allocated": 12.0, "ram_gib_allocated": 62.9, "host_cpu_threads": 48, "host_ram_gib": 251.5, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 32256 -r 3 -o json", "context_target": 32768, "context_note": "Prompt processing: 512 tokens timed after 32256 untimed tokens (context 32768 requested). Generation: 128 tokens timed after the same 32256 untimed tokens (context 32384 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 24, "repetitions": 3, "run": 2, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 32256, "context_requested": 32768, "mean_tps": 6231.08, "stdev_tps": 570.42, "samples_tps": [ 6218.6, 5666.99, 6807.64 ] }, "generation": { "timed_tokens": 128, "start_depth": 32256, "context_requested": 32384, "mean_tps": 142.61, "stdev_tps": 3.05, "samples_tps": [ 142.17, 139.81, 145.86 ] }, "gpu_memory_peak_mib": 18197, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 0, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T14:05:57Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T135058Z-4090-m47598/gemma-4-26b-a4b-32768-r2.json", "corrections": [], "page": "https://localrigcheck.com/gemma-4/" }, { "schema": 1, "id": "lab-2026-10-21t135058z-4090-m47598-gemma-4-26b-a4b-ud-q4-k-xl-32k-r3", "status": "current", "label": "RigCheck measurement", "identity": { "model": "gemma-4-26b-a4b", "repo": "unsloth/gemma-4-26B-A4B-it-GGUF", "revision": "c099eb48e663fd284577b04978a94ffccb261841", "file": "gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf", "quant": "UD-Q4_K_XL", "bytes": 17010980576, "sha256": "ef728c8e0c337fd1067b947af006e38a9ef2419e56feced4fd29b4bf0636e30c" }, "environment": { "provider": "Vast.ai", "machine_id": 47598, "host_id": 55116, "gpu": "NVIDIA GeForce RTX 4090", "gpu_memory_mib": 24564, "gpu_count": 1, "driver": "580.142", "cuda_driver_api": "13.0", "power_limit_w": 400.0, "power_limit_note": "host power limit 400 W; the card ships at 450 W", "pcie": { "gen_max": 4, "width_max": 16, "listed_bandwidth_gbps": 18.6 }, "cpu": "AMD EPYC 7402 24-Core Processor", "cpu_cores_allocated": 12.0, "ram_gib_allocated": 62.9, "host_cpu_threads": 48, "host_ram_gib": 251.5, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 32256 -r 3 -o json", "context_target": 32768, "context_note": "Prompt processing: 512 tokens timed after 32256 untimed tokens (context 32768 requested). Generation: 128 tokens timed after the same 32256 untimed tokens (context 32384 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 24, "repetitions": 3, "run": 3, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 32256, "context_requested": 32768, "mean_tps": 6232.57, "stdev_tps": 518.03, "samples_tps": [ 6341.13, 5668.86, 6687.72 ] }, "generation": { "timed_tokens": 128, "start_depth": 32256, "context_requested": 32384, "mean_tps": 146.01, "stdev_tps": 0.92, "samples_tps": [ 144.96, 146.67, 146.39 ] }, "gpu_memory_peak_mib": 18197, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 0, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T14:06:13Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T135058Z-4090-m47598/gemma-4-26b-a4b-32768-r3.json", "corrections": [], "page": "https://localrigcheck.com/gemma-4/" }, { "schema": 1, "id": "lab-2026-10-21t135058z-4090-m47598-gemma-4-26b-a4b-ud-q4-k-xl-4k-r1", "status": "current", "label": "RigCheck measurement", "identity": { "model": "gemma-4-26b-a4b", "repo": "unsloth/gemma-4-26B-A4B-it-GGUF", "revision": "c099eb48e663fd284577b04978a94ffccb261841", "file": "gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf", "quant": "UD-Q4_K_XL", "bytes": 17010980576, "sha256": "ef728c8e0c337fd1067b947af006e38a9ef2419e56feced4fd29b4bf0636e30c" }, "environment": { "provider": "Vast.ai", "machine_id": 47598, "host_id": 55116, "gpu": "NVIDIA GeForce RTX 4090", "gpu_memory_mib": 24564, "gpu_count": 1, "driver": "580.142", "cuda_driver_api": "13.0", "power_limit_w": 400.0, "power_limit_note": "host power limit 400 W; the card ships at 450 W", "pcie": { "gen_max": 4, "width_max": 16, "listed_bandwidth_gbps": 18.6 }, "cpu": "AMD EPYC 7402 24-Core Processor", "cpu_cores_allocated": 12.0, "ram_gib_allocated": 62.9, "host_cpu_threads": 48, "host_ram_gib": 251.5, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 3584 -r 3 -o json", "context_target": 4096, "context_note": "Prompt processing: 512 tokens timed after 3584 untimed tokens (context 4096 requested). Generation: 128 tokens timed after the same 3584 untimed tokens (context 3712 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 24, "repetitions": 3, "run": 1, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 3584, "context_requested": 4096, "mean_tps": 8909.66, "stdev_tps": 1001.33, "samples_tps": [ 9085.13, 7832.2, 9811.66 ] }, "generation": { "timed_tokens": 128, "start_depth": 3584, "context_requested": 3712, "mean_tps": 163.25, "stdev_tps": 1.14, "samples_tps": [ 161.99, 164.21, 163.57 ] }, "gpu_memory_peak_mib": 17597, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 0, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T14:05:01Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T135058Z-4090-m47598/gemma-4-26b-a4b-4096-r1.json", "corrections": [], "page": "https://localrigcheck.com/gemma-4/" }, { "schema": 1, "id": "lab-2026-10-21t135058z-4090-m47598-gemma-4-26b-a4b-ud-q4-k-xl-4k-r2", "status": "current", "label": "RigCheck measurement", "identity": { "model": "gemma-4-26b-a4b", "repo": "unsloth/gemma-4-26B-A4B-it-GGUF", "revision": "c099eb48e663fd284577b04978a94ffccb261841", "file": "gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf", "quant": "UD-Q4_K_XL", "bytes": 17010980576, "sha256": "ef728c8e0c337fd1067b947af006e38a9ef2419e56feced4fd29b4bf0636e30c" }, "environment": { "provider": "Vast.ai", "machine_id": 47598, "host_id": 55116, "gpu": "NVIDIA GeForce RTX 4090", "gpu_memory_mib": 24564, "gpu_count": 1, "driver": "580.142", "cuda_driver_api": "13.0", "power_limit_w": 400.0, "power_limit_note": "host power limit 400 W; the card ships at 450 W", "pcie": { "gen_max": 4, "width_max": 16, "listed_bandwidth_gbps": 18.6 }, "cpu": "AMD EPYC 7402 24-Core Processor", "cpu_cores_allocated": 12.0, "ram_gib_allocated": 62.9, "host_cpu_threads": 48, "host_ram_gib": 251.5, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 3584 -r 3 -o json", "context_target": 4096, "context_note": "Prompt processing: 512 tokens timed after 3584 untimed tokens (context 4096 requested). Generation: 128 tokens timed after the same 3584 untimed tokens (context 3712 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 24, "repetitions": 3, "run": 2, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 3584, "context_requested": 4096, "mean_tps": 8849.82, "stdev_tps": 994.3, "samples_tps": [ 9053.64, 7769.4, 9726.42 ] }, "generation": { "timed_tokens": 128, "start_depth": 3584, "context_requested": 3712, "mean_tps": 163.03, "stdev_tps": 1.23, "samples_tps": [ 161.61, 163.73, 163.74 ] }, "gpu_memory_peak_mib": 17609, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 0, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T14:05:12Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T135058Z-4090-m47598/gemma-4-26b-a4b-4096-r2.json", "corrections": [], "page": "https://localrigcheck.com/gemma-4/" }, { "schema": 1, "id": "lab-2026-10-21t135058z-4090-m47598-gemma-4-26b-a4b-ud-q4-k-xl-4k-r3", "status": "current", "label": "RigCheck measurement", "identity": { "model": "gemma-4-26b-a4b", "repo": "unsloth/gemma-4-26B-A4B-it-GGUF", "revision": "c099eb48e663fd284577b04978a94ffccb261841", "file": "gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf", "quant": "UD-Q4_K_XL", "bytes": 17010980576, "sha256": "ef728c8e0c337fd1067b947af006e38a9ef2419e56feced4fd29b4bf0636e30c" }, "environment": { "provider": "Vast.ai", "machine_id": 47598, "host_id": 55116, "gpu": "NVIDIA GeForce RTX 4090", "gpu_memory_mib": 24564, "gpu_count": 1, "driver": "580.142", "cuda_driver_api": "13.0", "power_limit_w": 400.0, "power_limit_note": "host power limit 400 W; the card ships at 450 W", "pcie": { "gen_max": 4, "width_max": 16, "listed_bandwidth_gbps": 18.6 }, "cpu": "AMD EPYC 7402 24-Core Processor", "cpu_cores_allocated": 12.0, "ram_gib_allocated": 62.9, "host_cpu_threads": 48, "host_ram_gib": 251.5, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 3584 -r 3 -o json", "context_target": 4096, "context_note": "Prompt processing: 512 tokens timed after 3584 untimed tokens (context 4096 requested). Generation: 128 tokens timed after the same 3584 untimed tokens (context 3712 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 24, "repetitions": 3, "run": 3, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 3584, "context_requested": 4096, "mean_tps": 8922.31, "stdev_tps": 1038.66, "samples_tps": [ 9153.61, 7787.49, 9825.82 ] }, "generation": { "timed_tokens": 128, "start_depth": 3584, "context_requested": 3712, "mean_tps": 163.35, "stdev_tps": 1.1, "samples_tps": [ 162.08, 164.13, 163.83 ] }, "gpu_memory_peak_mib": 17597, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 0, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T14:05:24Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T135058Z-4090-m47598/gemma-4-26b-a4b-4096-r3.json", "corrections": [], "page": "https://localrigcheck.com/gemma-4/" }, { "schema": 1, "id": "lab-2026-10-21t135058z-4090-m47598-qwen-3-8-27b-ud-q4-k-xl-32k-r1", "status": "current", "label": "RigCheck measurement", "identity": { "model": "qwen-3-8-27b", "repo": "unsloth/Qwen3.8-27B-GGUF", "revision": "4ca720788d1e01f1bff70c033e0d0028fd02e502", "file": "Qwen3.8-27B-UD-Q4_K_XL.gguf", "quant": "UD-Q4_K_XL", "bytes": 17559178144, "sha256": "3f227079003add2511437e5b1e94812e363385225bf6a9b47b0054a72bc8b01e" }, "environment": { "provider": "Vast.ai", "machine_id": 47598, "host_id": 55116, "gpu": "NVIDIA GeForce RTX 4090", "gpu_memory_mib": 24564, "gpu_count": 1, "driver": "580.142", "cuda_driver_api": "13.0", "power_limit_w": 400.0, "power_limit_note": "host power limit 400 W; the card ships at 450 W", "pcie": { "gen_max": 4, "width_max": 16, "listed_bandwidth_gbps": 18.6 }, "cpu": "AMD EPYC 7402 24-Core Processor", "cpu_cores_allocated": 12.0, "ram_gib_allocated": 62.9, "host_cpu_threads": 48, "host_ram_gib": 251.5, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m Qwen3.8-27B-UD-Q4_K_XL.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 32256 -r 3 -o json", "context_target": 32768, "context_note": "Prompt processing: 512 tokens timed after 32256 untimed tokens (context 32768 requested). Generation: 128 tokens timed after the same 32256 untimed tokens (context 32384 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 24, "repetitions": 3, "run": 1, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 32256, "context_requested": 32768, "mean_tps": 2418.66, "stdev_tps": 152.4, "samples_tps": [ 2483.64, 2244.54, 2527.79 ] }, "generation": { "timed_tokens": 128, "start_depth": 32256, "context_requested": 32384, "mean_tps": 43.57, "stdev_tps": 0.15, "samples_tps": [ 43.4, 43.65, 43.66 ] }, "gpu_memory_peak_mib": 18951, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 0, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T14:00:46Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T135058Z-4090-m47598/qwen-3-8-27b-32768-r1.json", "corrections": [], "page": "https://localrigcheck.com/qwen-3-8-27b/" }, { "schema": 1, "id": "lab-2026-10-21t135058z-4090-m47598-qwen-3-8-27b-ud-q4-k-xl-32k-r2", "status": "current", "label": "RigCheck measurement", "identity": { "model": "qwen-3-8-27b", "repo": "unsloth/Qwen3.8-27B-GGUF", "revision": "4ca720788d1e01f1bff70c033e0d0028fd02e502", "file": "Qwen3.8-27B-UD-Q4_K_XL.gguf", "quant": "UD-Q4_K_XL", "bytes": 17559178144, "sha256": "3f227079003add2511437e5b1e94812e363385225bf6a9b47b0054a72bc8b01e" }, "environment": { "provider": "Vast.ai", "machine_id": 47598, "host_id": 55116, "gpu": "NVIDIA GeForce RTX 4090", "gpu_memory_mib": 24564, "gpu_count": 1, "driver": "580.142", "cuda_driver_api": "13.0", "power_limit_w": 400.0, "power_limit_note": "host power limit 400 W; the card ships at 450 W", "pcie": { "gen_max": 4, "width_max": 16, "listed_bandwidth_gbps": 18.6 }, "cpu": "AMD EPYC 7402 24-Core Processor", "cpu_cores_allocated": 12.0, "ram_gib_allocated": 62.9, "host_cpu_threads": 48, "host_ram_gib": 251.5, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m Qwen3.8-27B-UD-Q4_K_XL.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 32256 -r 3 -o json", "context_target": 32768, "context_note": "Prompt processing: 512 tokens timed after 32256 untimed tokens (context 32768 requested). Generation: 128 tokens timed after the same 32256 untimed tokens (context 32384 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 24, "repetitions": 3, "run": 2, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 32256, "context_requested": 32768, "mean_tps": 2424.46, "stdev_tps": 134.14, "samples_tps": [ 2463.36, 2275.17, 2534.85 ] }, "generation": { "timed_tokens": 128, "start_depth": 32256, "context_requested": 32384, "mean_tps": 43.58, "stdev_tps": 0.1, "samples_tps": [ 43.47, 43.66, 43.62 ] }, "gpu_memory_peak_mib": 18951, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 0, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T14:01:19Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T135058Z-4090-m47598/qwen-3-8-27b-32768-r2.json", "corrections": [], "page": "https://localrigcheck.com/qwen-3-8-27b/" }, { "schema": 1, "id": "lab-2026-10-21t135058z-4090-m47598-qwen-3-8-27b-ud-q4-k-xl-32k-r3", "status": "current", "label": "RigCheck measurement", "identity": { "model": "qwen-3-8-27b", "repo": "unsloth/Qwen3.8-27B-GGUF", "revision": "4ca720788d1e01f1bff70c033e0d0028fd02e502", "file": "Qwen3.8-27B-UD-Q4_K_XL.gguf", "quant": "UD-Q4_K_XL", "bytes": 17559178144, "sha256": "3f227079003add2511437e5b1e94812e363385225bf6a9b47b0054a72bc8b01e" }, "environment": { "provider": "Vast.ai", "machine_id": 47598, "host_id": 55116, "gpu": "NVIDIA GeForce RTX 4090", "gpu_memory_mib": 24564, "gpu_count": 1, "driver": "580.142", "cuda_driver_api": "13.0", "power_limit_w": 400.0, "power_limit_note": "host power limit 400 W; the card ships at 450 W", "pcie": { "gen_max": 4, "width_max": 16, "listed_bandwidth_gbps": 18.6 }, "cpu": "AMD EPYC 7402 24-Core Processor", "cpu_cores_allocated": 12.0, "ram_gib_allocated": 62.9, "host_cpu_threads": 48, "host_ram_gib": 251.5, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m Qwen3.8-27B-UD-Q4_K_XL.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 32256 -r 3 -o json", "context_target": 32768, "context_note": "Prompt processing: 512 tokens timed after 32256 untimed tokens (context 32768 requested). Generation: 128 tokens timed after the same 32256 untimed tokens (context 32384 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 24, "repetitions": 3, "run": 3, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 32256, "context_requested": 32768, "mean_tps": 2408.87, "stdev_tps": 152.17, "samples_tps": [ 2471.88, 2235.31, 2519.42 ] }, "generation": { "timed_tokens": 128, "start_depth": 32256, "context_requested": 32384, "mean_tps": 43.59, "stdev_tps": 0.14, "samples_tps": [ 43.43, 43.67, 43.68 ] }, "gpu_memory_peak_mib": 18951, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 0, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T14:01:51Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T135058Z-4090-m47598/qwen-3-8-27b-32768-r3.json", "corrections": [], "page": "https://localrigcheck.com/qwen-3-8-27b/" }, { "schema": 1, "id": "lab-2026-10-21t135058z-4090-m47598-qwen-3-8-27b-ud-q4-k-xl-4k-r1", "status": "current", "label": "RigCheck measurement", "identity": { "model": "qwen-3-8-27b", "repo": "unsloth/Qwen3.8-27B-GGUF", "revision": "4ca720788d1e01f1bff70c033e0d0028fd02e502", "file": "Qwen3.8-27B-UD-Q4_K_XL.gguf", "quant": "UD-Q4_K_XL", "bytes": 17559178144, "sha256": "3f227079003add2511437e5b1e94812e363385225bf6a9b47b0054a72bc8b01e" }, "environment": { "provider": "Vast.ai", "machine_id": 47598, "host_id": 55116, "gpu": "NVIDIA GeForce RTX 4090", "gpu_memory_mib": 24564, "gpu_count": 1, "driver": "580.142", "cuda_driver_api": "13.0", "power_limit_w": 400.0, "power_limit_note": "host power limit 400 W; the card ships at 450 W", "pcie": { "gen_max": 4, "width_max": 16, "listed_bandwidth_gbps": 18.6 }, "cpu": "AMD EPYC 7402 24-Core Processor", "cpu_cores_allocated": 12.0, "ram_gib_allocated": 62.9, "host_cpu_threads": 48, "host_ram_gib": 251.5, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m Qwen3.8-27B-UD-Q4_K_XL.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 3584 -r 3 -o json", "context_target": 4096, "context_note": "Prompt processing: 512 tokens timed after 3584 untimed tokens (context 4096 requested). Generation: 128 tokens timed after the same 3584 untimed tokens (context 3712 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 24, "repetitions": 3, "run": 1, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 3584, "context_requested": 4096, "mean_tps": 3143.81, "stdev_tps": 239.97, "samples_tps": [ 3234.43, 2871.72, 3325.27 ] }, "generation": { "timed_tokens": 128, "start_depth": 3584, "context_requested": 3712, "mean_tps": 47.9, "stdev_tps": 0.17, "samples_tps": [ 47.7, 48.0, 48.0 ] }, "gpu_memory_peak_mib": 17131, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 0, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T13:59:39Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T135058Z-4090-m47598/qwen-3-8-27b-4096-r1.json", "corrections": [], "page": "https://localrigcheck.com/qwen-3-8-27b/" }, { "schema": 1, "id": "lab-2026-10-21t135058z-4090-m47598-qwen-3-8-27b-ud-q4-k-xl-4k-r2", "status": "current", "label": "RigCheck measurement", "identity": { "model": "qwen-3-8-27b", "repo": "unsloth/Qwen3.8-27B-GGUF", "revision": "4ca720788d1e01f1bff70c033e0d0028fd02e502", "file": "Qwen3.8-27B-UD-Q4_K_XL.gguf", "quant": "UD-Q4_K_XL", "bytes": 17559178144, "sha256": "3f227079003add2511437e5b1e94812e363385225bf6a9b47b0054a72bc8b01e" }, "environment": { "provider": "Vast.ai", "machine_id": 47598, "host_id": 55116, "gpu": "NVIDIA GeForce RTX 4090", "gpu_memory_mib": 24564, "gpu_count": 1, "driver": "580.142", "cuda_driver_api": "13.0", "power_limit_w": 400.0, "power_limit_note": "host power limit 400 W; the card ships at 450 W", "pcie": { "gen_max": 4, "width_max": 16, "listed_bandwidth_gbps": 18.6 }, "cpu": "AMD EPYC 7402 24-Core Processor", "cpu_cores_allocated": 12.0, "ram_gib_allocated": 62.9, "host_cpu_threads": 48, "host_ram_gib": 251.5, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m Qwen3.8-27B-UD-Q4_K_XL.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 3584 -r 3 -o json", "context_target": 4096, "context_note": "Prompt processing: 512 tokens timed after 3584 untimed tokens (context 4096 requested). Generation: 128 tokens timed after the same 3584 untimed tokens (context 3712 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 24, "repetitions": 3, "run": 2, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 3584, "context_requested": 4096, "mean_tps": 3159.33, "stdev_tps": 225.72, "samples_tps": [ 3260.65, 2900.71, 3316.63 ] }, "generation": { "timed_tokens": 128, "start_depth": 3584, "context_requested": 3712, "mean_tps": 47.85, "stdev_tps": 0.16, "samples_tps": [ 47.67, 47.96, 47.92 ] }, "gpu_memory_peak_mib": 17131, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 0, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T13:59:56Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T135058Z-4090-m47598/qwen-3-8-27b-4096-r2.json", "corrections": [], "page": "https://localrigcheck.com/qwen-3-8-27b/" }, { "schema": 1, "id": "lab-2026-10-21t135058z-4090-m47598-qwen-3-8-27b-ud-q4-k-xl-4k-r3", "status": "current", "label": "RigCheck measurement", "identity": { "model": "qwen-3-8-27b", "repo": "unsloth/Qwen3.8-27B-GGUF", "revision": "4ca720788d1e01f1bff70c033e0d0028fd02e502", "file": "Qwen3.8-27B-UD-Q4_K_XL.gguf", "quant": "UD-Q4_K_XL", "bytes": 17559178144, "sha256": "3f227079003add2511437e5b1e94812e363385225bf6a9b47b0054a72bc8b01e" }, "environment": { "provider": "Vast.ai", "machine_id": 47598, "host_id": 55116, "gpu": "NVIDIA GeForce RTX 4090", "gpu_memory_mib": 24564, "gpu_count": 1, "driver": "580.142", "cuda_driver_api": "13.0", "power_limit_w": 400.0, "power_limit_note": "host power limit 400 W; the card ships at 450 W", "pcie": { "gen_max": 4, "width_max": 16, "listed_bandwidth_gbps": 18.6 }, "cpu": "AMD EPYC 7402 24-Core Processor", "cpu_cores_allocated": 12.0, "ram_gib_allocated": 62.9, "host_cpu_threads": 48, "host_ram_gib": 251.5, "runtime": "llama.cpp", "build": 11783, "commit": "339f3242c", "backend": "CUDA", "image": "ghcr.io/ggml-org/llama.cpp:full-cuda-b11783" }, "workload": { "tool": "llama-bench", "command": "llama-bench -m Qwen3.8-27B-UD-Q4_K_XL.gguf -ngl 99 -fa 1 -ctk f16 -ctv f16 -p 512 -n 128 -d 3584 -r 3 -o json", "context_target": 4096, "context_note": "Prompt processing: 512 tokens timed after 3584 untimed tokens (context 4096 requested). Generation: 128 tokens timed after the same 3584 untimed tokens (context 3712 requested). The size llama.cpp actually allocated (it may round up) was not recorded.", "gpu_layers": "all (-ngl 99)", "flash_attention": true, "kv_cache": "f16", "batch": 2048, "ubatch": 512, "cpu_threads": 24, "repetitions": 3, "run": 3, "runs_in_cell": 3 }, "result": { "capacity": "fits", "supported": true, "loaded": true, "ran": true, "outcome": "ran", "prompt_processing": { "timed_tokens": 512, "start_depth": 3584, "context_requested": 4096, "mean_tps": 3152.13, "stdev_tps": 236.62, "samples_tps": [ 3208.2, 2892.5, 3355.68 ] }, "generation": { "timed_tokens": 128, "start_depth": 3584, "context_requested": 3712, "mean_tps": 47.87, "stdev_tps": 0.11, "samples_tps": [ 47.74, 47.92, 47.94 ] }, "gpu_memory_peak_mib": 17131, "gpu_memory_idle_mib": null, "gpu_memory_first_sample_mib": 0, "gpu_memory_note": "whole GPU memory.used, about every 200 ms; no idle value before loading was taken (first_sample is the first reading while llama-bench started)", "error": null }, "measured": "2026-10-21T14:00:14Z", "published": "2026-10-22", "valid_until_note": "Valid for this llama.cpp build and file revision; a new build or file can change speed.", "rights": { "license": "CC BY 4.0", "attribution": "RigCheck Lab, https://localrigcheck.com" }, "raw": "data/lab/raw/2026-10-21T135058Z-4090-m47598/qwen-3-8-27b-4096-r3.json", "corrections": [], "page": "https://localrigcheck.com/qwen-3-8-27b/" } ] }