# RigCheck > Free hardware checkers for running new open AI models locally. Numbers come from official repositories, model cards and published benchmarks; estimates are labeled. ## Pages - [Mistral Large 4 Hardware Requirements: ~1.05 TB FP8 Weights](https://localrigcheck.com/mistral-large-4/): How much memory does Mistral Large 4 need? 1.05T params (52B active): ~2.1 TB at BF16, ~1.05 TB at FP8, ~0.56–0.63 TB at 4-bit. GPU counts, Mac options, API cost. - [Gemma 4 Hardware Requirements: E4B 8 GB to 31B 24 GB VRAM](https://localrigcheck.com/gemma-4/): Gemma 4 locally: E4B fits an 8 GB GPU at 8-bit or a 16 GB Mac at 4-bit; the 12B wants 12 GB at 4-bit; 26B-A4B and 31B a 24 GB card or 32 GB Mac. - [LFM2.5-24B-A2B Hardware Requirements: 14.5 GB Q4, 24 GB GPU](https://localrigcheck.com/lfm-2-5-24b-a2b/): LFM2.5-24B-A2B locally: the official Q4_K_M GGUF (14.5 GB) needs about 17 GB at a 32K context, so a 24 GB GPU or a 32 GB Mac holds it. - [MiniMax H3 Hardware Requirements: 12 GB GPU, 480P Lab Run](https://localrigcheck.com/minimax-h3/): MiniMax H3 hardware routes: 480P FL2VA lab recipe with 12 GB NVIDIA VRAM, 32 GB RAM and NVMe; task downloads and the API-only full 2K limit. - [Qwen3.8-27B Hardware Requirements: 17.6 GB GGUF, 24 GB GPU](https://localrigcheck.com/qwen-3-8-27b/): Qwen3.8-27B locally: the 4-bit GGUF (17.6 GB) needs about 22 GB at a 32K context, so a 24 GB GPU or a 32 GB Mac holds it; 16 GB cards take 3-bit. - [Graphics Cards for Local AI: What Each GPU Can Run](https://localrigcheck.com/gpus/): What can your graphics card run locally? One page per GPU with every RigCheck checker's answer: RTX 3060, 4060 Ti, 3090, 4090, 5090. - [Macs for Local AI: What Each Memory Size Can Run](https://localrigcheck.com/macs/): What can your Mac run locally? One page per unified-memory size with every RigCheck checker's answer, from 24 GB to a 512 GB Mac Studio. - [VRAM for Local AI: What Each GPU Memory Size Can Run](https://localrigcheck.com/vram/): What can 8, 12, 16, 24 or 32 GB of VRAM run locally? One page per memory size with every RigCheck checker's answer, card by card. - [DeepSeek V4.1 Flash Hardware Requirements: 510.3 GB Weights](https://localrigcheck.com/deepseek-v4-1-flash/): DeepSeek V4.1 Flash locally: 510 GB official weights; DwarfStar streams the September weights on a 128 GB Mac or DGX Spark; 1017 GGUFs are experimental. - [Kolibri-1 Hardware Requirements: 78 GB FP8, 2× A100 80 GB](https://localrigcheck.com/kolibri-1/): Kolibri-1 hardware checker: 78B total versus 3.46B active, official FP8 server configurations, patched GGUF CPU/GPU route and context memory limits. - [Strata AI Engine Requirements: 32 GB RAM+VRAM, 8 GB GPU](https://localrigcheck.com/strata/): Strata AI engine requirements: 32 GB of RAM + VRAM together, an 8 GB NVIDIA or 12 GB AMD GPU, an SSD, or a Mac. Checker, speeds, LM Studio use. - [Nemotron 3.5 Hardware Requirements: 32 GB GPU for Bolt](https://localrigcheck.com/nemotron-3-5/): Nemotron 3.5 Bolt 74B: a 22 GB GGUF, but NVIDIA's llama.cpp recipes start at 32 GB GPUs or a 64 GB Spark. Super VL 120B: 64–192 GB. Files, Ollama, Mac. - [FLUX 3 Klein Hardware Requirements: 8 GB GPU (int8 ComfyUI)](https://localrigcheck.com/flux-3-image-klein/): FLUX 3 Image Klein locally: the int8 ComfyUI files run on 8 GB GPUs such as the RTX 3070 (BFL). File sizes, peak VRAM, the prompt upsampler and speeds. - [GLM-5.3-Flash Hardware Requirements: 100 GB for 1-Bit GGUF](https://localrigcheck.com/glm-5-3-flash/): GLM-5.3-Flash locally: about 100 GB of RAM + VRAM for 1-bit, 128–150 GB for 3-bit, 162–210 GB for 4-bit (Unsloth). What fits a PC, Mac or DGX Spark. - [Mistral Medium 3.5 Hardware Requirements: 75 GB 4-Bit Files](https://localrigcheck.com/mistral-medium-3-5/): Mistral Medium 3.5 (128B dense): 4-bit GGUF ~75 GB, 8-bit 133 GB, plus KV cache. What fits on RTX 3090s, 128 GB Macs, DGX Spark, and how fast. - [Qwen3.8-Flash-Next Hardware Requirements: 76 GB RAM + VRAM](https://localrigcheck.com/qwen-3-8-flash-next/): Qwen3.8-Flash-Next in llama.cpp: about 76 GB of RAM + VRAM for 1-bit, 86 GB for 3-bit, 115 GB for 4-bit. What fits your PC, Mac or DGX Spark. - [Qwen-Image-2.1 Hardware Requirements: 10.02 GB Q4 File Set](https://localrigcheck.com/qwen-image-2-1/): Qwen-Image-2.1 and Turbo hardware checker: full workflow files, GPU and RAM planning budgets, ComfyUI and GGUF routes, and research-license limits. - [About RigCheck](https://localrigcheck.com/about/): RigCheck builds free hardware checkers for new open AI models, using the developers' published requirements and benchmarks. - [New AI Model Alerts by Email](https://localrigcheck.com/alerts/): Get an email when RigCheck adds a hardware checker for a new open AI model, with links to the answers for your graphics card or Mac memory size. - [RigCheck Open Data: Model Fit and Hardware JSON (CC BY 4.0)](https://localrigcheck.com/data/): Download the data behind the RigCheck checkers as JSON: model file sizes, fit answers on common setups and the hardware list, free to reuse under CC BY 4.0. - [Local LLM Glossary: GGUF, Quantization, KV Cache, MoE](https://localrigcheck.com/glossary/): Plain-English meanings of the terms on RigCheck's checkers: GGUF and quant names like Q4_K_M or IQ2_XS, MoE experts, KV cache, offload, unified memory and tok/s. - [RigCheck Independence and Funding Statement](https://localrigcheck.com/independence/): Who pays for RigCheck, what it accepts, and the rule that money never changes an answer: no ads, no sponsors, labeled referral links only after a no. - [How RigCheck Checks Hardware: RigCheck Lab Method](https://localrigcheck.com/methodology/): How RigCheck Lab decides whether a GPU or Mac can run an open AI model: sources, memory math, labeled assumptions, verdicts, speeds and review. - [Privacy](https://localrigcheck.com/privacy/): RigCheck privacy policy: no accounts, checkers run in your browser, and Google Analytics and Microsoft Clarity only with your consent in Europe. - [Send Your Local AI Model Benchmark or Failure Report](https://localrigcheck.com/submit/): Ran an open AI model on your own GPU or Mac, or tried and it failed? Paste the llama-bench output or log; checked reports are published as community data. - [Local AI Models: Hardware Requirements for Every Model](https://localrigcheck.com/models/): Every open AI model RigCheck checks, by type: inputs, sizes, weights status, runtimes and license, with a hardware checker for each. - [RTX 3090 24 GB + 64 GB RAM: 7 of 16 Without Caveats](https://localrigcheck.com/gpus/rtx-3090/): What can an RTX 3090 (24 GB) run locally? Every RigCheck checker's answer with 32 or 64 GB of RAM: Qwen, Gemma 4, Mistral, Strata, image and video. - [RTX 4090 24 GB + 64 GB RAM: 7 of 16 Without Caveats](https://localrigcheck.com/gpus/rtx-4090/): What can an RTX 4090 (24 GB) run locally? Every RigCheck checker's answer with 32 or 64 GB of RAM: Qwen, Gemma 4, Mistral, Strata, image and video. - [RTX 5090 32 GB + 64 GB RAM: 7 of 16 Without Caveats](https://localrigcheck.com/gpus/rtx-5090/): What can an RTX 5090 (32 GB) run locally? Every RigCheck checker's answer with 32 or 64 GB of RAM: Qwen, Gemma 4, Mistral, Strata, image and video. - [RTX 4060 Ti 16 GB + 64 GB RAM: 7 of 15 Without Caveats](https://localrigcheck.com/gpus/rtx-4060-ti-16gb/): What can an RTX 4060 Ti 16 GB run locally? Every RigCheck checker's answer with 32 or 64 GB of RAM: Qwen, Gemma 4, Mistral, Strata, image and video. - [RTX 3060 12 GB + 64 GB RAM: 5 of 16 Without Caveats](https://localrigcheck.com/gpus/rtx-3060-12gb/): What can an RTX 3060 12 GB run locally? Every RigCheck checker's answer with 32 or 64 GB of RAM: Qwen, Gemma 4, Mistral, Strata, image and video. - [RTX 4070 Ti 12 GB + 64 GB RAM: 5 of 13 Without Caveats](https://localrigcheck.com/gpus/rtx-4070-ti/): What can an RTX 4070 Ti (12 GB) run locally? Every RigCheck checker's answer with 32 or 64 GB of RAM: Qwen, Gemma 4, Mistral, Strata, image and video. - [Radeon VII / MI50 16 GB + 64 GB RAM: 7 of 9 Answers](https://localrigcheck.com/gpus/radeon-vii-mi50/): What can a Radeon VII / Instinct MI50 16 GB run locally? Every RigCheck checker's answer with 32 or 64 GB of RAM: Qwen, Gemma 4, Mistral, Strata, image and video. - [Mac with 24 GB for Local AI: Answers for 8 of 10 Models](https://localrigcheck.com/macs/24gb/): What can a Mac with 24 GB run locally? Every RigCheck checker's answer for 24 GB of unified memory: Qwen, Gemma 4, Mistral, Strata and image models. - [Mac with 32 GB for Local AI: Answers for 9 of 11 Models](https://localrigcheck.com/macs/32gb/): What can a Mac with 32 GB run locally? Every RigCheck checker's answer for 32 GB of unified memory: Qwen, Gemma 4, Mistral, Strata and image models. - [Mac with 48 GB for Local AI: Answers for 9 of 11 Models](https://localrigcheck.com/macs/48gb/): What can a Mac with 48 GB run locally? Every RigCheck checker's answer for 48 GB of unified memory: Qwen, Gemma 4, Mistral, Strata and image models. - [Mac with 64 GB for Local AI: Answers for 10 of 11 Models](https://localrigcheck.com/macs/64gb/): What can a Mac with 64 GB run locally? Every RigCheck checker's answer for 64 GB of unified memory: Qwen, Gemma 4, Mistral, Strata and image models. - [Mac with 96 GB for Local AI: Answers for 10 of 12 Models](https://localrigcheck.com/macs/96gb/): What can a Mac with 96 GB run locally? Every RigCheck checker's answer for 96 GB of unified memory: Qwen, Gemma 4, Mistral, Strata and image models. - [Mac with M5 Max (128 GB): Answers for 15 of 16 Models](https://localrigcheck.com/macs/m5-max-128gb/): What can a Mac with M5 Max (128 GB) run locally? Every RigCheck checker's answer for 128 GB of unified memory: Qwen, Gemma 4, Mistral, Strata and image models. - [Mac Studio M3 Ultra (256 GB): Answers for 15 of 16 Models](https://localrigcheck.com/macs/m3-ultra-256gb/): What can a Mac Studio M3 Ultra (256 GB) run locally? Every RigCheck checker's answer for 256 GB of unified memory: Qwen, Gemma 4, Mistral, Strata and image models. - [Mac Studio M3 Ultra (512 GB): Answers for 7 of 8 Models](https://localrigcheck.com/macs/m3-ultra-512gb/): What can a Mac Studio M3 Ultra (512 GB) run locally? Every RigCheck checker's answer for 512 GB of unified memory: Qwen, Gemma 4, Mistral, Strata and image models. - [8 GB VRAM: What Local AI Models Can An 8 GB GPU Run?](https://localrigcheck.com/vram/8gb/): What can an 8 GB graphics card run locally? Every RigCheck checker's answer for each of 20 cards with 8 GB of VRAM: Qwen, Gemma 4, Strata, image and video. - [12 GB VRAM: What Local AI Models Can A 12 GB GPU Run?](https://localrigcheck.com/vram/12gb/): What can a 12 GB graphics card run locally? Every RigCheck checker's answer for each of 12 cards with 12 GB of VRAM: Qwen, Gemma 4, Strata, image and video. - [16 GB VRAM: What Local AI Models Can A 16 GB GPU Run?](https://localrigcheck.com/vram/16gb/): What can a 16 GB graphics card run locally? Every RigCheck checker's answer for each of 21 cards with 16 GB of VRAM: Qwen, Gemma 4, Strata, image and video. - [24 GB VRAM: What Local AI Models Can A 24 GB GPU Run?](https://localrigcheck.com/vram/24gb/): What can a 24 GB graphics card run locally? Every RigCheck checker's answer for each of 7 cards with 24 GB of VRAM: Qwen, Gemma 4, Strata, image and video. - [32 GB VRAM: What Local AI Models Can A 32 GB GPU Run?](https://localrigcheck.com/vram/32gb/): What can a 32 GB graphics card run locally? Every RigCheck checker's answer for each of 3 cards with 32 GB of VRAM: Qwen, Gemma 4, Strata, image and video. ## Key facts (as of each page's update date) - Short answer: With GeForce RTX 3060 (12 GB) + 32 GB RAM, Gemma 4 12B (Q5_K_M), Gemma 4 E2B (BF16), Gemma 4 E4B (Q8_0), Qwen3.8-27B (UD-IQ2_XXS) fit with every layer in GPU memory; with GeForce RTX 3090 / 3090 Ti (24 GB) + 64 GB RAM, Gemma 4 12B (Q8_0), Gemma 4 26B-A4B (UD-Q4_K_XL), Gemma 4 31B (IQ4_XS), Gemma 4 E2B (BF16), Gemma 4 E4B (BF16), Qwen3.8-27B (UD-Q4_K_XL) fit with every layer in GPU memory. These are RigCheck capacity estimates for llama.cpp at a 32K context in the finder's quick mode, not measured setups; larger models need system RAM or a Mac, and image and video models are judged by planning budgets and published recipes. Every model: the model list. (kind of answer: RigCheck capacity estimates in the finder's quick mode (32K context, llama.cpp); updated 2026-10-22, source page https://localrigcheck.com/) - Short answer: the weights alone take about 2.1 TB at BF16, 1.05 TB at FP8, ~0.56–0.63 TB at 4-bit and ~0.28–0.51 TB at 2–3-bit GGUF quants. Only 52B parameters are active per token (49B not counting embeddings and output layers), but all 1.05T must be held in memory. By capacity alone that is an 8-GPU server at FP8, or a 512 GB Mac Studio for the smallest quants; it will not fit an RTX 4090/5090. These are memory estimates; runtime support is unknown until the weights ship. (kind of answer: capacity estimate; updated 2026-10-19, source page https://localrigcheck.com/mistral-large-4/) - Short answer: at a 32K context with llama.cpp (RigCheck estimates), Gemma 4 E4B at 8-bit (Q8_0) fits an 8 GB card with about 4.2 GB of it in system RAM (its 3.7 GB of embedding tables and 0.5 GB of host buffers), and a 16 GB Mac holds its Q4_K_M; E2B fits either at 8-bit or more. Gemma 4 12B at 4-bit (UD-Q4_K_XL) needs about 11 GB, fitting a 12 GB card; an 8 GB card holds the 3-bit UD-IQ3_XXS. 26B-A4B at 4-bit (UD-Q4_K_XL) needs about 21 GB and 31B about 23 GB with IQ4_XS (16.4 GB): both fit 24 GB cards (RTX 3090/4090, RX 7900 XTX) or 32 GB Macs, while the 31B's Q4_K_M fits 32 GB cards or 48 GB Macs but exceeds a 32 GB Mac's 24 GB usable budget. 16 GB cards take their 3-bit or 2-bit files. (kind of answer: capacity estimate; routes named: llama.cpp (community; official); updated 2026-10-22, source page https://localrigcheck.com/gemma-4/) - Short answer: with llama.cpp the official 4-bit Q4_K_M (14.5 GB) needs about 17 GB in total at a 32K context (RigCheck estimate). It fits a 24 GB card (RTX 3090, 4090) or a 32 GB Mac. A 16 GB card holds Unsloth's 3-bit UD-Q3_K_XL fully at 32K, or Q4_K_M with 7 of its 40 layers in system RAM (slower). Without a GPU, 32 GB of RAM holds Q6_K. Ollama's lfm2.5:24b is a 15 GB download. (kind of answer: ; routes named: llama.cpp (official), Ollama (official); updated 2026-10-22, source page https://localrigcheck.com/lfm-2-5-24b-a2b/) - Short answer: the BF16 weights for one task are 144 GB, so no consumer GPU holds them: every published consumer recipe streams the weights from system RAM, or NVMe when RAM is short. The smallest published setup is an NVIDIA GPU with 12 GB of VRAM and 32 GB of RAM plus an NVMe drive at 480P (SGLang, run with the memory capped; a real 32 GB PC will be slower). Keeping every weight on the GPU takes 4× H100 80 GB (SGLang), two 80 GB cards, or two 48 GB cards with int8 (diffusers), or one 192 GB AMD MI300X. We found no published Mac run. License: MiniMax's open-weights license does not cover the US, the EU, the UK or South Korea. (kind of answer: published recipe match; routes named: SGLang (official), Diffusers (official); updated 2026-10-22, source page https://localrigcheck.com/minimax-h3/) - Short answer: with llama.cpp the 4-bit UD-Q4_K_XL (17.6 GB) needs about 22 GB in total at a 32K context (RigCheck estimate). It fits a 24 GB card (RTX 3090, 4090, RX 7900 XTX) or a 32 GB Mac. A 16 GB card holds the 3-bit UD-IQ3_XXS at 32K, or the 4-bit file partly in system RAM (slower). An 8 GB card holds most layers of the 1-bit UD-IQ1_S, the rest in system RAM; with the 4-bit file most layers run from system RAM. Unsloth's total-memory table (RAM + VRAM) gives 16–19 GB for 4-bit and 12–14 GB for 3-bit. (kind of answer: capacity estimate; routes named: llama.cpp (community); updated 2026-10-22, source page https://localrigcheck.com/qwen-3-8-27b/) - Short answer: you need about 128 GB of memory; no consumer GPU on its own has a published route. The official weights are 510.3 GB (FP8 with FP4 experts). DwarfStar, a community engine, streams its 365.7 GB Q2 file of the September weights from SSD on one 128 GB Mac or DGX Spark. For the 1017 update every path is experimental: DwarfStar published a 1017 Q2 file on October 19 that its guides do not cover yet, and Unsloth's smallest 1017 GGUF is 233.0 GB, which Unsloth estimates needs 128–145 GB of RAM plus VRAM and which loads only in its unmerged llama.cpp branch or Unsloth Desktop. Ollama and LM Studio offer the model only in their clouds. (kind of answer: published route match; routes named: DwarfStar (community), llama.cpp (experimental), Ollama (cloud-only); updated 2026-10-21, source page https://localrigcheck.com/deepseek-v4-1-flash/) - Short answer: Kolibri has 78.10B total parameters and 3.46B active per token. The official model card gives an approximately 78 GB FP8 weight footprint; its 32 downloadable shards total 78.84 GB before runtime/cache overhead. Official minimum hardware: 2× A100 80 GB, 2× H100 SXM5, or 1× H200/B200/B300; the official routes are Aleph Alpha's vLLM plugin and, since October 20, 2026, Transformers 5.20 or newer. An independent 47.45 GB Q4_K_M GGUF has a community CPU / RTX 5070 + 128 GB RAM recipe, but requires a custom llama.cpp patch. Stock Ollama support and Apple Metal execution are not verified. (kind of answer: official route; routes named: vLLM (official), Transformers (official), llama.cpp (patched); updated 2026-10-21, source page https://localrigcheck.com/kolibri-1/) - Short answer: at minimum, 32 GB of RAM and graphics memory (VRAM) together (64 GB of RAM runs every standard size with all experts in RAM), an NVIDIA GeForce RTX 20–50 series with 8 GB of VRAM (slow; 12 GB recommended) or a supported AMD Radeon with 12 GB+, ~80 GB free on an SSD (more in low-RAM mode), and Windows 10/11 or Linux. Apple Silicon Macs (M1 or newer, macOS 15 or newer) run it through Strata's Metal engine, ready-made in its Mac app: 64 GB of unified memory keeps the fastest size (Q2_0) in memory, and 32–48 GB Macs stream the experts from the SSD. (kind of answer: Strata's published requirements; routes named: Strata (official; experimental); updated 2026-10-21, source page https://localrigcheck.com/strata/) - Short answer: Bolt's official GGUF is 22.0 GB (2-bit), but NVIDIA's own llama.cpp recipes start at a 32 GB GPU (RTX 5090) or a 64 GB RTX Spark, with 64K context; NVIDIA's 256K-context recipe is for a 48 GB GPU or a 128 GB Spark. Unsloth budgets about 32 GB of RAM or unified memory for short contexts. NVIDIA publishes no recipe for 24 GB cards such as the RTX 4090. Super VL does not fit one consumer GPU: Unsloth's GGUFs need about 64 GB of total memory (RAM + VRAM, or unified memory) for 1-bit, 96 GB for NVFP4 and 128 GB for UD-Q4_K_XL (4-bit). (kind of answer: published route match; routes named: llama.cpp (official; community); updated 2026-10-20, source page https://localrigcheck.com/nemotron-3-5/) - Short answer: Black Forest Labs says the int8 files in ComfyUI run on 8 GB GPUs such as the RTX 3070. BFL does not say whether its 8 GB figure includes the 9B upsampler it says not to skip, so by our reading on 8 GB expect to run it separately or switch it off (with weaker results). BFL's own code has no int8 path, and its NVFP4 path needs a Blackwell GPU, which by our reading includes the RTX 50 series. (kind of answer: published route match; routes named: ComfyUI (official); updated 2026-10-19, source page https://localrigcheck.com/flux-3-image-klein/) - Short answer: with llama.cpp and Unsloth's GGUF files you need about 100 GB of memory in total (system RAM plus VRAM, or unified memory) for 1-bit, 115 GB for 2-bit, 128–150 GB for 3-bit and 162–210 GB for 4-bit, by Unsloth's table. No single consumer GPU holds even the smallest file (93.1 GB); on a PC most of the model sits in system RAM. Unsloth says its 3-bit UD-IQ3_XXS fits 128 GB devices such as a Mac or DGX Spark; that is its claim, not a published run, and leaves little room for context (a Mac also needs its GPU wired-memory limit raised, per the MLX LM docs). The official FP8 weights are 328.3 GB (306 GiB): KTransformers runs them with one RTX 40 or 50 card (its tutorial gives no VRAM minimum), at least 350 GB of free system memory and a CPU with AVX-512; vLLM and SGLang target server GPUs (H100 and newer, AMD Instinct MI350X/MI355X). (kind of answer: published route match; routes named: llama.cpp (community), KTransformers (community), vLLM (official), SGLang (official); updated 2026-10-19, source page https://localrigcheck.com/glm-5-3-flash/) - Short answer: the 4-bit GGUF files are about 75 GB, 8-bit 133 GB, BF16 250 GB and the smallest 2-bit file 35 GB; add 11.8 GB of KV cache at a 32K context (16-bit). By our estimate at 32K, UD-Q4_K_XL fits fully on a 128 GB Mac or DGX Spark, and one 96 GB RTX PRO 6000 holds IQ4_XS. With 64 GB of RAM, a single RTX 4090 runs Q3_K_M and a single RTX 5090 runs IQ4_XS only with offload: in each case about 54 GB including input embeddings, host buffers, layers and their KV cache sit in system RAM. With two RTX 3090s, UD-Q4_K_XL puts about 49 GB including input embeddings, host buffers, layers and KV cache in RAM. That caps speed at roughly 2 tokens per second (RigCheck ceiling at 90 GB/s dual-channel DDR5, short context). (kind of answer: capacity estimate; updated 2026-10-19, source page https://localrigcheck.com/mistral-medium-3-5/) - Short answer: with llama.cpp you need about 76 GB of memory in total (system RAM plus VRAM, or unified memory) for the smallest 1-bit file, about 86 GB for 3-bit (UD-IQ3_XXS) and 115 GB for the 4-bit UD-Q4_K_XL at a 32K context (RigCheck estimate; Unsloth lists 75 GB for 1-bit and 96–114 GB for 4-bit). Even 1-bit is 72.5 GB: the 51B-parameter n-gram table stays at 4 bits or more, and llama.cpp keeps that table on the CPU side, never on the GPU. A GPU is optional: by our estimate a PC with no GPU and 96 GB of RAM holds UD-IQ3_XXS; an RTX 4090 or 5090 with 128 GB of RAM holds UD-Q4_K_XL with most of it in RAM; one 96 GB RTX PRO 6000 with 64 GB of RAM holds UD-Q4_K_XL with CPU inputs and host buffers in RAM. A 128 GB Mac or DGX Spark just holds UD-Q4_K_XL at 32K without image input (a Ryzen AI Max+ 395 too, if about 85 GB is GPU-accessible and the rest stays with the CPU; untested); the 8-bit file needs a 256 GB Mac. The KV cache is small: 1.0 GB at 32K. With 12–16 GB of VRAM and 32–64 GB of RAM, use the Strata engine instead. (kind of answer: capacity estimate; routes named: llama.cpp (community), Strata (community); updated 2026-10-19, source page https://localrigcheck.com/qwen-3-8-flash-next/) - Short answer: ComfyUI BF16 files total 32.44 GB; INT8 ConvRot files total 17.28 GB, including the encoder and VAE. Base GGUF Q4 plus encoder and VAE is 10.02 GB, or 11.19 GB with the separate vision projector for editing. These are file totals, not VRAM minima: attention, reference inputs, prefix caches and VAE decoding add memory. Turbo's eight steps do not shrink its weights. Base CUDA ComfyUI and official Diffusers routes are documented; Turbo ComfyUI sampling remains unverified here. Base GGUF has a stable-diffusion.cpp route. Mac memory budgets remain conditional; no verified device measurement here. The Qwen Research License does not grant commercial use; commercial generation services need separate permission. (kind of answer: planning budget; routes named: ComfyUI (community), Diffusers (official), stable-diffusion.cpp (community); updated 2026-10-19, source page https://localrigcheck.com/qwen-image-2-1/) - With 64 GB of system RAM, the checkers have an answer for an RTX 3090 (24 GB of VRAM) for 14 of the 16 models, sizes and routes they cover: 7 without caveats (Qwen3.8-Flash-Next in the Strata engine, Qwen3.8-27B in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B); Qwen3.8-Flash-Next in llama.cpp, Mistral Medium 3.5 in llama.cpp, Qwen-Image-2.1 (image generation), MiniMax H3 (video + audio), Nemotron 3.5 Bolt 74B-A9B, Nemotron 3.5 Super VL 120B-A12B, FLUX 3 Image Klein (int8, ComfyUI) with caveats (part in system RAM, an experimental path, a planning budget or an unmeasured peak). With 32 GB, 6 without caveats (Qwen3.8-27B in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B). The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups. (kind of answer: each model checker's own quick-mode answer: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/gpus/rtx-3090/) - With 64 GB of system RAM, the checkers have an answer for an RTX 4090 (24 GB of VRAM) for 14 of the 16 models, sizes and routes they cover: 7 without caveats (Qwen3.8-Flash-Next in the Strata engine, Qwen3.8-27B in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B); Qwen3.8-Flash-Next in llama.cpp, Mistral Medium 3.5 in llama.cpp, Qwen-Image-2.1 (image generation), MiniMax H3 (video + audio), Nemotron 3.5 Bolt 74B-A9B, Nemotron 3.5 Super VL 120B-A12B, FLUX 3 Image Klein (int8, ComfyUI) with caveats (part in system RAM, an experimental path, a planning budget or an unmeasured peak). With 32 GB, 6 without caveats (Qwen3.8-27B in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B). The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups. (kind of answer: each model checker's own quick-mode answer: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/gpus/rtx-4090/) - With 64 GB of system RAM, the checkers have an answer for an RTX 5090 (32 GB of VRAM) for 14 of the 16 models, sizes and routes they cover: 7 without caveats (Qwen3.8-Flash-Next in the Strata engine, Qwen3.8-27B in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B); Qwen3.8-Flash-Next in llama.cpp, Mistral Medium 3.5 in llama.cpp, Qwen-Image-2.1 (image generation), MiniMax H3 (video + audio), Nemotron 3.5 Bolt 74B-A9B, Nemotron 3.5 Super VL 120B-A12B, FLUX 3 Image Klein (int8, ComfyUI) with caveats (part in system RAM, an experimental path, a planning budget or an unmeasured peak). With 32 GB, 6 without caveats (Qwen3.8-27B in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B). The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups. (kind of answer: each model checker's own quick-mode answer: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/gpus/rtx-5090/) - With 64 GB of system RAM, the checkers have an answer for an RTX 4060 Ti 16 GB (16 GB of VRAM) for 12 of the 15 models, sizes and routes they cover: 7 without caveats (Qwen3.8-Flash-Next in the Strata engine, Qwen3.8-27B in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B); Mistral Medium 3.5 in llama.cpp, MiniMax H3 (video + audio), Nemotron 3.5 Bolt 74B-A9B, Nemotron 3.5 Super VL 120B-A12B, FLUX 3 Image Klein (int8, ComfyUI) with caveats (part in system RAM, an experimental path, a planning budget or an unmeasured peak). With 32 GB, 6 without caveats (Qwen3.8-27B in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B). The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups. (kind of answer: each model checker's own quick-mode answer: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/gpus/rtx-4060-ti-16gb/) - With 64 GB of system RAM, the checkers have an answer for an RTX 3060 12 GB (12 GB of VRAM) for 13 of the 16 models, sizes and routes they cover: 5 without caveats (Qwen3.8-Flash-Next in the Strata engine, Qwen3.8-27B in llama.cpp, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B); Mistral Medium 3.5 in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Qwen-Image-2.1 (image generation), MiniMax H3 (video + audio), Nemotron 3.5 Bolt 74B-A9B, Nemotron 3.5 Super VL 120B-A12B, FLUX 3 Image Klein (int8, ComfyUI) with caveats (part in system RAM, an experimental path, a planning budget or an unmeasured peak). With 32 GB, 4 without caveats (Qwen3.8-27B in llama.cpp, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B). The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups. (kind of answer: each model checker's own quick-mode answer: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/gpus/rtx-3060-12gb/) - With 64 GB of system RAM, the checkers have an answer for an RTX 4070 Ti (12 GB of VRAM) for 10 of the 13 models, sizes and routes they cover: 5 without caveats (Qwen3.8-Flash-Next in the Strata engine, Qwen3.8-27B in llama.cpp, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B); Gemma 4 31B, Gemma 4 26B-A4B, Nemotron 3.5 Bolt 74B-A9B, Nemotron 3.5 Super VL 120B-A12B, FLUX 3 Image Klein (int8, ComfyUI) with caveats (part in system RAM, an experimental path, a planning budget or an unmeasured peak). With 32 GB, 4 without caveats (Qwen3.8-27B in llama.cpp, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B). The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups. (kind of answer: each model checker's own quick-mode answer: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/gpus/rtx-4070-ti/) - With 64 GB of system RAM, the checkers have an answer for a Radeon VII / Instinct MI50 16 GB (16 GB of VRAM) for 7 of the 9 models, sizes and routes they cover: Qwen3.8-Flash-Next in the Strata engine, Qwen3.8-27B in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B with caveats (part in system RAM, an experimental path, a planning budget or an unmeasured peak). With 32 GB, 0 without caveats. The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups. (kind of answer: each model checker's own quick-mode answer: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/gpus/radeon-vii-mi50/) - For a Mac with 24 GB (24 GB of unified memory), the checkers have an answer for 8 of the 10 models, sizes and routes they cover: Qwen3.8-27B in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B once you raise the Mac's GPU wired-memory limit; Qwen3.8-Flash-Next in the Strata engine, Qwen-Image-2.1 (image generation) only with caveats (an experimental path, a planning budget or an unmeasured peak). The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups; how fast a model answers depends on the chip's memory bandwidth. (kind of answer: each model checker's own quick-mode answer: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/macs/24gb/) - For a Mac with 32 GB (32 GB of unified memory), the checkers have an answer for 9 of the 11 models, sizes and routes they cover: Qwen3.8-27B in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B once you raise the Mac's GPU wired-memory limit; Qwen3.8-Flash-Next in the Strata engine, Qwen-Image-2.1 (image generation), Nemotron 3.5 Bolt 74B-A9B only with caveats (an experimental path, a planning budget or an unmeasured peak). The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups; how fast a model answers depends on the chip's memory bandwidth. (kind of answer: each model checker's own quick-mode answer: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/macs/32gb/) - For a Mac with 48 GB (48 GB of unified memory), the checkers have an answer for 9 of the 11 models, sizes and routes they cover: Qwen3.8-27B in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B once you raise the Mac's GPU wired-memory limit; Qwen3.8-Flash-Next in the Strata engine, Qwen-Image-2.1 (image generation), Nemotron 3.5 Bolt 74B-A9B only with caveats (an experimental path, a planning budget or an unmeasured peak). The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups; how fast a model answers depends on the chip's memory bandwidth. (kind of answer: each model checker's own quick-mode answer: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/macs/48gb/) - For a Mac with 64 GB (64 GB of unified memory), the checkers have an answer for 10 of the 11 models, sizes and routes they cover: Qwen3.8-27B in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B once you raise the Mac's GPU wired-memory limit; Qwen3.8-Flash-Next in the Strata engine, Qwen-Image-2.1 (image generation), Nemotron 3.5 Bolt 74B-A9B, Nemotron 3.5 Super VL 120B-A12B only with caveats (an experimental path, a planning budget or an unmeasured peak). The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups; how fast a model answers depends on the chip's memory bandwidth. (kind of answer: each model checker's own quick-mode answer: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/macs/64gb/) - For a Mac with 96 GB (96 GB of unified memory), the checkers have an answer for 10 of the 12 models, sizes and routes they cover: Qwen3.8-27B in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B once you raise the Mac's GPU wired-memory limit; Qwen3.8-Flash-Next in the Strata engine, Qwen-Image-2.1 (image generation), Nemotron 3.5 Bolt 74B-A9B, Nemotron 3.5 Super VL 120B-A12B only with caveats (an experimental path, a planning budget or an unmeasured peak). The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups; how fast a model answers depends on the chip's memory bandwidth. (kind of answer: each model checker's own quick-mode answer: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/macs/96gb/) - For a Mac with M5 Max (128 GB) (128 GB of unified memory, 614 GB/s), the checkers have an answer for 15 of the 16 models, sizes and routes they cover: Qwen3.8-Flash-Next in llama.cpp, Mistral Medium 3.5 in llama.cpp, Qwen3.8-27B in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B once you raise the Mac's GPU wired-memory limit; Qwen3.8-Flash-Next in the Strata engine, Qwen-Image-2.1 (image generation), Nemotron 3.5 Bolt 74B-A9B, Nemotron 3.5 Super VL 120B-A12B, GLM-5.3-Flash (Unsloth GGUF), DeepSeek V4.1 Flash September release, DeepSeek V4.1 Flash 1017 update only with caveats (an experimental path, a planning budget or an unmeasured peak). The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups; how fast a model answers depends on the chip's memory bandwidth. (kind of answer: each model checker's own quick-mode answer: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/macs/m5-max-128gb/) - For a Mac Studio M3 Ultra (256 GB) (256 GB of unified memory, 819 GB/s), the checkers have an answer for 15 of the 16 models, sizes and routes they cover: Qwen3.8-Flash-Next in llama.cpp, Mistral Medium 3.5 in llama.cpp, Qwen3.8-27B in llama.cpp, Gemma 4 31B, Gemma 4 26B-A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B once you raise the Mac's GPU wired-memory limit; Qwen3.8-Flash-Next in the Strata engine, Qwen-Image-2.1 (image generation), Nemotron 3.5 Bolt 74B-A9B, Nemotron 3.5 Super VL 120B-A12B, GLM-5.3-Flash (Unsloth GGUF), DeepSeek V4.1 Flash September release, DeepSeek V4.1 Flash 1017 update only with caveats (an experimental path, a planning budget or an unmeasured peak). The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups; how fast a model answers depends on the chip's memory bandwidth. (kind of answer: each model checker's own quick-mode answer: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/macs/m3-ultra-256gb/) - For a Mac Studio M3 Ultra (512 GB) (512 GB of unified memory, 819 GB/s), the checkers have an answer for 7 of the 8 models, sizes and routes they cover: Qwen3.8-Flash-Next in llama.cpp, Mistral Medium 3.5 in llama.cpp once you raise the Mac's GPU wired-memory limit; Qwen3.8-Flash-Next in the Strata engine, Qwen-Image-2.1 (image generation), GLM-5.3-Flash (Unsloth GGUF), DeepSeek V4.1 Flash September release, DeepSeek V4.1 Flash 1017 update only with caveats (an experimental path, a planning budget or an unmeasured peak). The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups; how fast a model answers depends on the chip's memory bandwidth. (kind of answer: each model checker's own quick-mode answer: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/macs/m3-ultra-512gb/) - With 64 GB of system RAM, across the 20 graphics cards with 8 GB of VRAM in RigCheck's finder, the checkers have an answer for 11 of the 14 models, sizes and routes they cover on at least one of them. Placed wholly or partly on the card, by RigCheck's llama.cpp capacity estimates: Qwen3.8-27B in llama.cpp (10 of 20), Gemma 4 31B (10 of 20), Gemma 4 26B-A4B (10 of 20), Gemma 4 12B (10 of 20, without caveats or conditions on 8), Gemma 4 E4B (10 of 20, without caveats or conditions on 8), Gemma 4 E2B (10 of 20, without caveats or conditions on 8). On the processor in system RAM alone, the card unused: Qwen3.8-27B in llama.cpp (10 of 20), Gemma 4 31B (10 of 20), Gemma 4 26B-A4B (10 of 20), Gemma 4 12B (10 of 20), Gemma 4 E4B (10 of 20), Gemma 4 E2B (10 of 20). Answers that name no placement (Strata's published requirements, published route and recipe matches, planning budgets; some rest on system RAM alone, as the table says): Qwen3.8-Flash-Next in the Strata engine (18 of 20), Qwen-Image-2.1 (image generation) (2 of 20), Nemotron 3.5 Bolt 74B-A9B (10 of 20), Nemotron 3.5 Super VL 120B-A12B (10 of 20), FLUX 3 Image Klein (int8, ComfyUI) (9 of 20). Which cards get an answer, and which answer, depends on the card and not only on its memory: the table lists them card by card. The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups. (kind of answer: each model checker's own quick-mode answer for every card with this memory: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/vram/8gb/) - With 64 GB of system RAM, across the 12 graphics cards with 12 GB of VRAM in RigCheck's finder, the checkers have an answer for 13 of the 16 models, sizes and routes they cover on at least one of them. Placed wholly or partly on the card, by RigCheck's llama.cpp capacity estimates: Mistral Medium 3.5 in llama.cpp (1 of 12), Qwen3.8-27B in llama.cpp (9 of 12, without caveats or conditions on 8), Gemma 4 31B (9 of 12), Gemma 4 26B-A4B (9 of 12), Gemma 4 12B (9 of 12, without caveats or conditions on 8), Gemma 4 E4B (9 of 12, without caveats or conditions on 8), Gemma 4 E2B (9 of 12, without caveats or conditions on 8). On the processor in system RAM alone, the card unused: Qwen3.8-27B in llama.cpp (3 of 12), Gemma 4 31B (3 of 12), Gemma 4 26B-A4B (3 of 12), Gemma 4 12B (3 of 12), Gemma 4 E4B (3 of 12), Gemma 4 E2B (3 of 12). Answers that name no placement (Strata's published requirements, published route and recipe matches, planning budgets; some rest on system RAM alone, as the table says): Qwen3.8-Flash-Next in the Strata engine (all 12), Qwen-Image-2.1 (image generation) (2 of 12), MiniMax H3 (video + audio) (4 of 12), Nemotron 3.5 Bolt 74B-A9B (9 of 12), Nemotron 3.5 Super VL 120B-A12B (9 of 12), FLUX 3 Image Klein (int8, ComfyUI) (8 of 12). Which cards get an answer, and which answer, depends on the card and not only on its memory: the table lists them card by card. The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups. (kind of answer: each model checker's own quick-mode answer for every card with this memory: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/vram/12gb/) - With 64 GB of system RAM, across the 21 graphics cards with 16 GB of VRAM in RigCheck's finder, the checkers have an answer for 13 of the 16 models, sizes and routes they cover on at least one of them. Placed wholly or partly on the card, by RigCheck's llama.cpp capacity estimates: Mistral Medium 3.5 in llama.cpp (4 of 21), Qwen3.8-27B in llama.cpp (17 of 21, without caveats or conditions on 8), Gemma 4 31B (17 of 21, without caveats or conditions on 8), Gemma 4 26B-A4B (17 of 21, without caveats or conditions on 8), Gemma 4 12B (17 of 21, without caveats or conditions on 8), Gemma 4 E4B (17 of 21, without caveats or conditions on 8), Gemma 4 E2B (17 of 21, without caveats or conditions on 8). On the processor in system RAM alone, the card unused: Qwen3.8-27B in llama.cpp (4 of 21), Gemma 4 31B (4 of 21), Gemma 4 26B-A4B (4 of 21), Gemma 4 12B (4 of 21), Gemma 4 E4B (4 of 21), Gemma 4 E2B (4 of 21). Answers that name no placement (Strata's published requirements, published route and recipe matches, planning budgets; some rest on system RAM alone, as the table says): Qwen3.8-Flash-Next in the Strata engine (all 21), Qwen-Image-2.1 (image generation) (3 of 21), MiniMax H3 (video + audio) (7 of 21), Nemotron 3.5 Bolt 74B-A9B (17 of 21), Nemotron 3.5 Super VL 120B-A12B (17 of 21), FLUX 3 Image Klein (int8, ComfyUI) (8 of 21). Which cards get an answer, and which answer, depends on the card and not only on its memory: the table lists them card by card. The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups. (kind of answer: each model checker's own quick-mode answer for every card with this memory: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/vram/16gb/) - With 64 GB of system RAM, across the 7 graphics cards with 24 GB of VRAM in RigCheck's finder, the checkers have an answer for 14 of the 16 models, sizes and routes they cover on at least one of them. Placed wholly or partly on the card, by RigCheck's llama.cpp capacity estimates: Qwen3.8-Flash-Next in llama.cpp (3 of 7), Mistral Medium 3.5 in llama.cpp (3 of 7), Qwen3.8-27B in llama.cpp (5 of 7, without caveats or conditions on 4), Gemma 4 31B (5 of 7, without caveats or conditions on 4), Gemma 4 26B-A4B (5 of 7, without caveats or conditions on 4), Gemma 4 12B (5 of 7, without caveats or conditions on 4), Gemma 4 E4B (5 of 7, without caveats or conditions on 4), Gemma 4 E2B (5 of 7, without caveats or conditions on 4). On the processor in system RAM alone, the card unused: Qwen3.8-27B in llama.cpp (2 of 7), Gemma 4 31B (2 of 7), Gemma 4 26B-A4B (2 of 7), Gemma 4 12B (2 of 7), Gemma 4 E4B (2 of 7), Gemma 4 E2B (2 of 7). Answers that name no placement (Strata's published requirements, published route and recipe matches, planning budgets; some rest on system RAM alone, as the table says): Qwen3.8-Flash-Next in the Strata engine (all 7), Qwen-Image-2.1 (image generation) (3 of 7), MiniMax H3 (video + audio) (3 of 7), Nemotron 3.5 Bolt 74B-A9B (5 of 7), Nemotron 3.5 Super VL 120B-A12B (5 of 7), FLUX 3 Image Klein (int8, ComfyUI) (4 of 7). Which cards get an answer, and which answer, depends on the card and not only on its memory: the table lists them card by card. The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups. (kind of answer: each model checker's own quick-mode answer for every card with this memory: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/vram/24gb/) - With 64 GB of system RAM, across the 3 graphics cards with 32 GB of VRAM in RigCheck's finder, the checkers have an answer for 14 of the 16 models, sizes and routes they cover on at least one of them. Placed wholly or partly on the card, by RigCheck's llama.cpp capacity estimates: Qwen3.8-Flash-Next in llama.cpp (1 of 3), Mistral Medium 3.5 in llama.cpp (1 of 3), Qwen3.8-27B in llama.cpp (2 of 3, without caveats or conditions on 1), Gemma 4 31B (2 of 3, without caveats or conditions on 1), Gemma 4 26B-A4B (2 of 3, without caveats or conditions on 1), Gemma 4 12B (2 of 3, without caveats or conditions on 1), Gemma 4 E4B (2 of 3, without caveats or conditions on 1), Gemma 4 E2B (2 of 3, without caveats or conditions on 1). On the processor in system RAM alone, the card unused: Qwen3.8-27B in llama.cpp (1 of 3), Gemma 4 31B (1 of 3), Gemma 4 26B-A4B (1 of 3), Gemma 4 12B (1 of 3), Gemma 4 E4B (1 of 3), Gemma 4 E2B (1 of 3). Answers that name no placement (Strata's published requirements, published route and recipe matches, planning budgets; some rest on system RAM alone, as the table says): Qwen3.8-Flash-Next in the Strata engine (all 3), Qwen-Image-2.1 (image generation) (1 of 3), MiniMax H3 (video + audio) (1 of 3), Nemotron 3.5 Bolt 74B-A9B (2 of 3), Nemotron 3.5 Super VL 120B-A12B (2 of 3), FLUX 3 Image Klein (int8, ComfyUI) (1 of 3). Which cards get an answer, and which answer, depends on the card and not only on its memory: the table lists them card by card. The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups. (kind of answer: each model checker's own quick-mode answer for every card with this memory: capacity estimates, planning budgets or published recipes; updated 2026-10-22, source page https://localrigcheck.com/vram/32gb/) ## Data - https://localrigcheck.com/data/: what each file holds, the CC BY 4.0 terms and how to credit RigCheck; changes are mirrored at https://github.com/Qtguns/rigcheck-data. - https://localrigcheck.com/data/models.json, https://localrigcheck.com/data/strata.json, https://localrigcheck.com/data/hardware.json: the checkers' file sizes, benchmarks (with their source and type) and hardware list as JSON, CC BY 4.0, attribution "RigCheck, https://localrigcheck.com". - https://localrigcheck.com/data/fit/deepseek-v4-1-flash.json: DeepSeek V4.1 Flash, the checker's answers on 8 common setups (the page's "Common setups" table) as JSON, CC BY 4.0. - https://localrigcheck.com/data/fit/flux-3-image-klein.json: FLUX 3 Image Klein, the checker's answers on 8 common setups (the page's "Common setups" table) as JSON, CC BY 4.0. - https://localrigcheck.com/data/fit/gemma-4.json: Gemma 4, the checker's answers on 8 common setups (the page's "Common setups" table) as JSON, CC BY 4.0. - https://localrigcheck.com/data/fit/glm-5-3-flash.json: GLM-5.3-Flash, the checker's answers on 8 common setups (the page's "Common setups" table) as JSON, CC BY 4.0. - https://localrigcheck.com/data/fit/kolibri-1.json: Kolibri-1, the checker's answers on 8 common setups (the page's "Common setups" table) as JSON, CC BY 4.0. - https://localrigcheck.com/data/fit/minimax-h3.json: MiniMax H3, the checker's answers on 8 common setups (the page's "Common setups" table) as JSON, CC BY 4.0. - https://localrigcheck.com/data/fit/mistral-medium-3-5.json: Mistral Medium 3.5, the checker's answers on 8 common setups (the page's "Common setups" table) as JSON, CC BY 4.0. - https://localrigcheck.com/data/fit/nemotron-3-5.json: Nemotron 3.5, the checker's answers on 8 common setups (the page's "Common setups" table) as JSON, CC BY 4.0. - https://localrigcheck.com/data/fit/qwen-3-8-27b.json: Qwen3.8-27B, the checker's answers on 8 common setups (the page's "Common setups" table) as JSON, CC BY 4.0. - https://localrigcheck.com/data/fit/qwen-3-8-flash-next.json: Qwen3.8-Flash-Next, the checker's answers on 8 common setups (the page's "Common setups" table) as JSON, CC BY 4.0. - https://localrigcheck.com/data/fit/qwen-image-2-1.json: Qwen-Image-2.1, the checker's answers on 8 common setups (the page's "Common setups" table) as JSON, CC BY 4.0. - https://localrigcheck.com/data/fit/strata.json: Strata checker, the checker's answers on 8 common setups (the page's "Common setups" table) as JSON, CC BY 4.0. - https://localrigcheck.com/data/lab.json: RigCheck Lab's own measurements (24 llama-bench runs, each with its card, host, file, context, samples and correction history) as JSON, CC BY 4.0; shown on the model and graphics card pages they cover, under "RigCheck Lab measurements". ## Citing RigCheck - Can I run model X, which file or quantization fits my card or Mac, which runtime supports it: the model's page. Each has a short answer near the top, sourced tables, a FAQ and its dates. Example: https://localrigcheck.com/qwen-3-8-27b/ - What can graphics card X run: its /gpus// page, every model's answer for that card in the checker's own words. Example: https://localrigcheck.com/gpus/rtx-3090/ - What fits on a Mac with N GB of unified memory: its /macs// page. Example: https://localrigcheck.com/macs/24gb/ - What fits in N GB of graphics memory, card by card: its /vram/gb/ page. Example: https://localrigcheck.com/vram/8gb/ - Which open models RigCheck covers, with type, sizes and first setup: the model list, https://localrigcheck.com/models/ - Machine-readable data: https://localrigcheck.com/data/models.json, https://localrigcheck.com/data/hardware.json and https://localrigcheck.com/data/catalog.json (CC BY 4.0). Attribute as "RigCheck, https://localrigcheck.com" and link the page the number comes from. - Link the public page and quote its "Page updated" or "Sources checked" date. Interactive checkers run on https://tools.localrigcheck.com inside those pages; that origin is not indexed, so do not cite its URLs. ## Recent updates - 2026-10-22: Gemma 4: RigCheck Lab measurements of Gemma 4 26B-A4B: UD-IQ3_XXS on an RTX 4060 Ti 16 GB and UD-Q4_K_XL on an RTX 4090, at about 4K and 32K context, three runs each (https://localrigcheck.com/gemma-4/) - 2026-10-22: LFM2.5-24B-A2B: Checker launched: will LFM2.5-24B-A2B fit your GPU, RAM or Mac, from the official and Unsloth GGUF file sizes, with the command to run it (https://localrigcheck.com/lfm-2-5-24b-a2b/) - 2026-10-22: MiniMax H3: vLLM's recipe now lists single-card RTX 4090 / RTX 5090 / DGX Spark configurations for FastH3, FastVideo's few-step derivative: noted as configuration guidance the recipe has not verified on a GPU, on a vLLM-Omni pull-request branch (#9167), not matched by the checker (https://localrigcheck.com/minimax-h3/) - 2026-10-22: Qwen3.8-27B: RigCheck Lab measurements: UD-IQ3_XXS on an RTX 4060 Ti 16 GB and UD-Q4_K_XL on an RTX 4090, at about 4K and 32K context, three runs each (https://localrigcheck.com/qwen-3-8-27b/) - 2026-10-21: DeepSeek V4.1 Flash: DeepInfer, the newly open-sourced framework behind DeepSeek's own services, added to the runtimes: it runs V4.1 Flash (1017 by default) only on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs under Linux, so it is no route for Macs, and in RigCheck's reading of NVIDIA's list none for gaming cards or the DGX Spark (https://localrigcheck.com/deepseek-v4-1-flash/) - 2026-10-21: Kolibri-1: Official route update: Aleph Alpha's model card now also documents Transformers, which supports Kolibri natively from version 5.20; the card gives no separate hardware for it, and the checker has no Transformers route (https://localrigcheck.com/kolibri-1/) - 2026-10-21: Strata checker: Strata 0.2.1 docs re-checked: memory now counts RAM and VRAM together (32 GB minimum, sizes picked by RAM + VRAM, low-RAM mode wherever RAM is short), Gyro-S and Gyro-M sizes added, ready-made apps for Windows, Linux and Mac, AMD packages carry picture readers on Windows too (untested there), the Coder's authors updated its first shard on October 17 (95.7% of SWE-bench Verified), but Strata 0.2.1 still installs the September 29 build (91%) (https://localrigcheck.com/strata/) - 2026-10-20: DeepSeek V4.1 Flash: 1017 update: experimental 128 GB paths from DwarfStar's new 1017 Q2 file (365.7 GB, not yet in its guides) and Unsloth's 1017 GGUF from 233.0 GB (unmerged llama.cpp branch); DwarfStar on ROCm (Strix Halo) added as experimental; download commands fixed (https://localrigcheck.com/deepseek-v4-1-flash/) - 2026-10-20: Nemotron 3.5: Super VL: ggml-org re-uploaded its GGUFs; Q4_0 is now 71.9 GB (was 71.8 GB), NVFP4 unchanged at 76.9 GB (https://localrigcheck.com/nemotron-3-5/) - 2026-10-19: Gemma 4: Jetson Orin Nano: NVIDIA's Jetson AI Lab route for Gemma 4 E2B (Q4_K_S in llama.cpp), with the larger sizes NVIDIA points to bigger Jetsons (https://localrigcheck.com/gemma-4/)