What can an RTX 5090 run locally?
- Page updated
- Answers from each model's own checker
- Card memory from NVIDIA
What runs on an RTX 5090? Every answer below is the one RigCheck's own checker for that model gives for this card, with 32 GB or 64 GB of system RAM. Nothing on this page is judged separately.
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.
RTX 5090: model by model
Language models
| Model (checker) | With 32 GB RAM | With 64 GB RAM |
|---|---|---|
| Qwen3.8-Flash-Next in the Strata engine | ! Runs, with caveats IQ2_XS No measured speed for this hardware and setup yet. Runs in low-RAM mode: the 24 GB+ card holds part of the experts and the rest are copied from a file on the SSD into RAM at start. Measured with 64 GB of RAM; with 32 GB your PC runs it in low-RAM mode, which is slower. For code, the Coder (IQ1_M) runs here with its experts copied into RAM at start; Strata: it is weaker outside code and in languages other than English. Strata's installer pre-selects IQ3_S for 64 GB of RAM + VRAM, which would read experts from the SSD in low-RAM mode here (slower); RigCheck picks IQ2_XS, whose experts Strata copies into RAM at start on this PC. | ✓ Runs IQ3_S No measured speed for this hardware and setup yet. Fits by the docs' rule (experts + ~10 GB), with little room to spare: close the browser while it runs. |
| Qwen3.8-Flash-Next in llama.cpp | ✕ Does not fit Even the smallest file (UD-IQ1_S, 72.5 GB) needs about 76.1 GB. | ! Fits with caveats (capacity estimate) UD-Q2_K_XL · GPU and system RAM (53.1 GB in RAM) No measured speed for this hardware and setup yet. Flash-Next uses combined memory totals; per-card layer placement is not modeled. Close fits are uncertain until runtime allocation is checked. RigCheck capacity estimate, not a measured setup. |
| Mistral Medium 3.5 in llama.cpp | ! Fits with caveats (capacity estimate) UD-IQ2_XXS · GPU and system RAM (52 of 88 layers on the GPU): a dense model runs slowly from RAM No measured speed for this hardware and setup yet. RigCheck capacity estimate, not a measured setup. | ! Fits with caveats (capacity estimate) IQ4_XS · GPU and system RAM (28 of 88 layers on the GPU): a dense model runs slowly from RAM No measured speed for this hardware and setup yet. RigCheck capacity estimate, not a measured setup. |
| Qwen3.8-27B in llama.cpp | ✓ Fits (capacity estimate) UD-Q6_K_XL · GPU memory (input embeddings in system RAM) No measured speed for this hardware and setup yet. RigCheck capacity estimate, not a measured setup. | ✓ Fits (capacity estimate) UD-Q6_K_XL · GPU memory (input embeddings in system RAM) No measured speed for this hardware and setup yet. RigCheck capacity estimate, not a measured setup. |
| Gemma 4 (E2B, E4B, 12B, 26B-A4B, 31B) in llama.cpp | ✓ Best size: fits (capacity estimate) Gemma 4 31B: fits (capacity estimate) (Q5_K_M, GPU memory (input embeddings in system RAM)); Gemma 4 26B-A4B: fits (capacity estimate) (Q8_0, GPU memory (input embeddings in system RAM)); Gemma 4 12B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); Gemma 4 E4B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); Gemma 4 E2B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)) No measured speed for this hardware and setup yet. RigCheck capacity estimate, not a measured setup. | ✓ Best size: fits (capacity estimate) Gemma 4 31B: fits (capacity estimate) (Q5_K_M, GPU memory (input embeddings in system RAM)); Gemma 4 26B-A4B: fits (capacity estimate) (Q8_0, GPU memory (input embeddings in system RAM)); Gemma 4 12B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); Gemma 4 E4B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); Gemma 4 E2B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)) No measured speed for this hardware and setup yet. RigCheck capacity estimate, not a measured setup. |
| Nemotron 3.5 (Bolt 74B-A9B, Super VL 120B-A12B) | ! Best size: a published route or recipe matches, with caveats Nemotron 3.5 Bolt 74B-A9B: a published route or recipe matches, with caveats (Matches NVIDIA's 28GB llama.cpp recipe: 64K context, q8_0 KV cache, one request at a time); Nemotron 3.5 Super VL 120B-A12B: a published route or recipe matches, with caveats (Matches Unsloth's 64 GB budget: UD-IQ1_M (46.1 GB) or UD-IQ2_XXS (51.1 GB)) Nemotron 3.5 Bolt 74B-A9B: NVIDIA's 256K-context recipe needs an RTX Spark 128 GB or a 48 GB+ discrete GPU; this hardware gets the 65,536-token recipe. Build the latest llama.cpp from source first. No speed is published for this machine; NVIDIA's 84 tokens per second is from a 128 GB RTX Spark. Nemotron 3.5 Super VL 120B-A12B: Unsloth's planning estimate for a modest context, including the OS, runtime, KV cache and vision file; not a measured run. This setup counts 32 GB of VRAM + 32 GB of RAM = 64 GB. Longer contexts and video need more. | ! Best size: a published route or recipe matches, with caveats Nemotron 3.5 Bolt 74B-A9B: a published route or recipe matches, with caveats (Matches NVIDIA's 28GB llama.cpp recipe: 64K context, q8_0 KV cache, one request at a time); Nemotron 3.5 Super VL 120B-A12B: a published route or recipe matches, with caveats (Matches Unsloth's 96 GB budget: UD-Q3_K_XL (65.8 GB) or NVFP4 (79 GB)) Nemotron 3.5 Bolt 74B-A9B: NVIDIA's 256K-context recipe needs an RTX Spark 128 GB or a 48 GB+ discrete GPU; this hardware gets the 65,536-token recipe. Build the latest llama.cpp from source first. No speed is published for this machine; NVIDIA's 84 tokens per second is from a 128 GB RTX Spark. Nemotron 3.5 Super VL 120B-A12B: Unsloth's planning estimate for a modest context, including the OS, runtime, KV cache and vision file; not a measured run. This setup counts 32 GB of VRAM + 64 GB of RAM = 96 GB. Longer contexts and video need more. |
No answer here: Mistral Large 4, Kolibri-1 (official FP8 route), GLM-5.3-Flash (Unsloth GGUF), DeepSeek V4.1 Flash (September release, 1017 update) in DwarfStar
| Model (checker) | With 32 GB RAM | With 64 GB RAM |
|---|---|---|
| Mistral Large 4 | … Weights not released smallest estimate ~210 GB vs 64 GB of memory (VRAM + RAM, or a Mac's unified memory) | … Weights not released smallest estimate ~210 GB vs 96 GB of memory (VRAM + RAM, or a Mac's unified memory) |
| Kolibri-1 (official FP8 route) | – Not covered | – Not covered |
| GLM-5.3-Flash (Unsloth GGUF) | ✕ Does not fit Unsloth's table puts the 1-bit GGUF (93.1–97.6 GB) at 100 GB of total memory, RAM + VRAM or unified memory. This setup counts 32 GB of VRAM + 32 GB of RAM = 64 GB. | ✕ Does not fit Unsloth's table puts the 1-bit GGUF (93.1–97.6 GB) at 100 GB of total memory, RAM + VRAM or unified memory. This setup counts 32 GB of VRAM + 64 GB of RAM = 96 GB. |
| DeepSeek V4.1 Flash (September release, 1017 update) in DwarfStar | – Best size: not covered DeepSeek V4.1 Flash September release: not covered; DeepSeek V4.1 Flash 1017 update: not covered | – Best size: not covered DeepSeek V4.1 Flash September release: not covered; DeepSeek V4.1 Flash 1017 update: not covered |
Image models
| Model (checker) | With 32 GB RAM | With 64 GB RAM |
|---|---|---|
| Qwen-Image-2.1 (image generation) | ! Planning estimate only Within the all-components-resident planning budget; allocation unverified Uncalibrated allowances, not a minimum or a guarantee. Full checkpoint files are counted in host RAM even when mmap or sequential loading can use less. Actual peak depends on attention, prefix caches, VAE tiling, reference sizes and backend. No speed prediction. | ! Planning estimate only Within the all-components-resident planning budget; allocation unverified Uncalibrated allowances, not a minimum or a guarantee. Full checkpoint files are counted in host RAM even when mmap or sequential loading can use less. Actual peak depends on attention, prefix caches, VAE tiling, reference sizes and backend. No speed prediction. |
| FLUX 3 Image Klein (int8, ComfyUI) | ! A published route or recipe matches, with caveats Matches BFL's 8 GB guidance: the int8 files in ComfyUI BFL: the int8 ComfyUI files run on 8 GB GPUs such as the RTX 3070. Image model, int8 text encoder and VAE together are 11.0 GB; by our reading, on a card smaller than that ComfyUI keeps them on the GPU one at a time and needs system RAM for the rest. BFL gives no RAM figure. The prompt upsampler BFL says not to skip is a 9B model (6.72 GB w4a8, 9.75 GB int8 or 18.8 GB BF16 in ComfyUI); BFL does not say whether its 8 GB figure includes it. When it does not fit with the image model, run it on its own first (BFL's README) or switch it off in the ComfyUI template, with weaker results. NVIDIA measured about 4 seconds per 1 MP image on an RTX 5090 (distilled, int8, 16 steps, ComfyUI, upsampler not included). | ! A published route or recipe matches, with caveats Matches BFL's 8 GB guidance: the int8 files in ComfyUI BFL: the int8 ComfyUI files run on 8 GB GPUs such as the RTX 3070. Image model, int8 text encoder and VAE together are 11.0 GB; by our reading, on a card smaller than that ComfyUI keeps them on the GPU one at a time and needs system RAM for the rest. BFL gives no RAM figure. The prompt upsampler BFL says not to skip is a 9B model (6.72 GB w4a8, 9.75 GB int8 or 18.8 GB BF16 in ComfyUI); BFL does not say whether its 8 GB figure includes it. When it does not fit with the image model, run it on its own first (BFL's README) or switch it off in the ComfyUI template, with weaker results. NVIDIA measured about 4 seconds per 1 MP image on an RTX 5090 (distilled, int8, 16 steps, ComfyUI, upsampler not included). |
Video models
| Model (checker) | With 32 GB RAM | With 64 GB RAM |
|---|---|---|
| MiniMax H3 (video + audio) | ! A published route or recipe matches, with caveats Hardware memory matches a published video recipe; exact clip peak unverified H3-Base FL2VA, source recipe canvases only. Disk-streaming routes assume an NVMe drive. Host-RAM and lab assumptions remain in the full checker; full 2K needs an API. Matching recipes: SGLang, one 24 GB card (measured): one card, DiT and text encoder streamed layer by layer; GPU peak about 18 GB; RTX 4090 D: 1344×768, 107 frames, 20 steps in 405.6 s at BF16, 303.3 s with convrot_int8 (host RAM not stated for this run; we require the 32 GB of the cookbook's smallest recipe (RigCheck assumption); a 4090 on a 32 GB host has not been measured) | SGLang Recipe A (lab): one card, DiT, text encoder and VAE streamed from the checkpoint file on an NVMe drive (Recipe A), BF16; measured at 864×480 (at the native 768P the cookbook says to use video_vae=24 if the decode runs out of memory, not measured); lab run with VRAM capped at 12 GiB and a simulated 32 GB host: shows the memory fits, not your card's speed; a real 32 GB machine re-reads the weights from disk on every step | ! A published route or recipe matches, with caveats Hardware memory matches a published video recipe; exact clip peak unverified H3-Base FL2VA, source recipe canvases only. Disk-streaming routes assume an NVMe drive. Host-RAM and lab assumptions remain in the full checker; full 2K needs an API. Matching recipes: SGLang, one 24 GB card (measured): one card, DiT and text encoder streamed layer by layer; GPU peak about 18 GB; RTX 4090 D: 1344×768, 107 frames, 20 steps in 405.6 s at BF16, 303.3 s with convrot_int8 (host RAM not stated for this run; we require the 32 GB of the cookbook's smallest recipe (RigCheck assumption); a 4090 on a 32 GB host has not been measured) | SGLang Recipe A (lab): one card, DiT, text encoder and VAE streamed from the checkpoint file on an NVMe drive (Recipe A), BF16; measured at 864×480 (at the native 768P the cookbook says to use video_vae=24 if the decode runs out of memory, not measured); lab run with VRAM capped at 12 GiB and a simulated 32 GB host: shows the memory fits, not your card's speed; a real 32 GB machine re-reads the weights from disk on every step |
Quick setup, as in the home page finder: one RTX 5090, a 32K context and no image input for the llama.cpp capacity estimates (Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash are matched against published recipes, memory tables and routes at their own context instead (NVIDIA's 64K / 256K recipes; Unsloth's GLM table states no context length; DwarfStar's routes per DeepSeek checkpoint (documented for the September weights, experimental for 1017), not llama.cpp).), the better of Windows and Linux for Strata, an NVMe SSD; for the image and video rows, Qwen-Image (a planning budget): base model, INT8, 512×512, batch one, no reference images; Kolibri-1: the official FP8 route at a 32K context; MiniMax H3: local H3-Base FL2VA recipes, assuming an NVMe drive; Nemotron 3.5: Bolt and Super VL on one card with the RAM given; FLUX 3 Image Klein: Black Forest Labs' int8 ComfyUI route on the card. "With caveats" means part of the model sits in system RAM (slower), an unmeasured peak, or a route the checker marks conditional. Speeds appear only where they were measured on this very card in this setup; official estimates and runs on other setups are left out (the full checkers show them, labeled). Pick this card in a checker for your exact setup.
Measured a local model on your own hardware, or tried and it failed? Send us the numbers: paste the benchmark output or log into the form (or email it). After checking it we add it to this page, labeled as a community report and shared under CC BY 4.0.
About the RTX 5090
RTX 5090: 32 GB GDDR7 of video memory (NVIDIA GeForce RTX 5090 specifications). For the large mixture-of-experts models here, system RAM matters as much as VRAM: the parts that do not fit on the card are kept in RAM.
FAQ
Can an RTX 5090 run Qwen3.8-Flash-Next in the Strata engine?
Yes, with 64 GB of system RAM in RigCheck's quick setup: IQ3_S; No measured speed for this hardware and setup yet; Fits by the docs' rule (experts + ~10 GB), with little room to spare: close the browser while it runs. The full checker takes your exact setup (RAM, disk, operating system, processor) and gives the install steps.
Can an RTX 5090 run Qwen3.8-Flash-Next in llama.cpp?
With caveats, by RigCheck's capacity estimate, with 64 GB of system RAM in RigCheck's quick setup: UD-Q2_K_XL · GPU and system RAM (53.1 GB in RAM); No measured speed for this hardware and setup yet; Flash-Next uses combined memory totals; per-card layer placement is not modeled. Close fits are uncertain until runtime allocation is checked; RigCheck capacity estimate, not a measured setup. The full checker takes your exact setup (context length, RAM, image input) and gives the command to start it.
Can an RTX 5090 run Mistral Medium 3.5 in llama.cpp?
With caveats, by RigCheck's capacity estimate, with 64 GB of system RAM in RigCheck's quick setup: IQ4_XS · GPU and system RAM (28 of 88 layers on the GPU): a dense model runs slowly from RAM; No measured speed for this hardware and setup yet; RigCheck capacity estimate, not a measured setup. The full checker takes your exact setup (context length, RAM, image input) and gives the command to start it.
Can an RTX 5090 run Qwen3.8-27B in llama.cpp?
Yes, by RigCheck's capacity estimate, with 64 GB of system RAM in RigCheck's quick setup: UD-Q6_K_XL · GPU memory (input embeddings in system RAM); No measured speed for this hardware and setup yet; RigCheck capacity estimate, not a measured setup. The full checker takes your exact setup (context length, RAM, image input) and gives the command to start it.
Can an RTX 5090 run Gemma 4 (E2B, E4B, 12B, 26B-A4B, 31B) in llama.cpp?
Yes, by RigCheck's capacity estimate, with 64 GB of system RAM in RigCheck's quick setup: Gemma 4 31B: fits (capacity estimate) (Q5_K_M, GPU memory (input embeddings in system RAM)); Gemma 4 26B-A4B: fits (capacity estimate) (Q8_0, GPU memory (input embeddings in system RAM)); Gemma 4 12B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); Gemma 4 E4B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); Gemma 4 E2B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); No measured speed for this hardware and setup yet; RigCheck capacity estimate, not a measured setup. The full checker takes your exact setup (context length, RAM, image input) and gives the command to start it.
Can an RTX 5090 run Qwen-Image-2.1 (image generation)?
Only as a planning estimate: the checker's uncalibrated memory budget, offload peak unverified, with 64 GB of system RAM in RigCheck's quick setup: Within the all-components-resident planning budget; allocation unverified; Uncalibrated allowances, not a minimum or a guarantee. Full checkpoint files are counted in host RAM even when mmap or sequential loading can use less. Actual peak depends on attention, prefix caches, VAE tiling, reference sizes and backend. No speed prediction. The full checker lets you change the task, resolution, references and offload (batch is fixed at one), and lists the files.
Can an RTX 5090 run MiniMax H3 (video + audio)?
A published route or recipe matches, with caveats, with 64 GB of system RAM in RigCheck's quick setup: Hardware memory matches a published video recipe; exact clip peak unverified; H3-Base FL2VA, source recipe canvases only. Disk-streaming routes assume an NVMe drive. Host-RAM and lab assumptions remain in the full checker; full 2K needs an API; Matching recipes: SGLang, one 24 GB card (measured): one card, DiT and text encoder streamed layer by layer; GPU peak about 18 GB; RTX 4090 D: 1344×768, 107 frames, 20 steps in 405.6 s at BF16, 303.3 s with convrot_int8 (host RAM not stated for this run; we require the 32 GB of the cookbook's smallest recipe (RigCheck assumption); a 4090 on a 32 GB host has not been measured) | SGLang Recipe A (lab): one card, DiT, text encoder and VAE streamed from the checkpoint file on an NVMe drive (Recipe A), BF16; measured at 864×480 (at the native 768P the cookbook says to use video_vae=24 if the decode runs out of memory, not measured); lab run with VRAM capped at 12 GiB and a simulated 32 GB host: shows the memory fits, not your card's speed; a real 32 GB machine re-reads the weights from disk on every step. The full checker lets you change the route or recipe settings and lists the files and the published recipe.
Can an RTX 5090 run Nemotron 3.5 (Bolt 74B-A9B, Super VL 120B-A12B)?
A published route or recipe matches, with caveats, with 64 GB of system RAM in RigCheck's quick setup: Nemotron 3.5 Bolt 74B-A9B: a published route or recipe matches, with caveats (Matches NVIDIA's 28GB llama.cpp recipe: 64K context, q8_0 KV cache, one request at a time); Nemotron 3.5 Super VL 120B-A12B: a published route or recipe matches, with caveats (Matches Unsloth's 96 GB budget: UD-Q3_K_XL (65.8 GB) or NVFP4 (79 GB)); Nemotron 3.5 Bolt 74B-A9B: NVIDIA's 256K-context recipe needs an RTX Spark 128 GB or a 48 GB+ discrete GPU; this hardware gets the 65,536-token recipe. Build the latest llama.cpp from source first. No speed is published for this machine; NVIDIA's 84 tokens per second is from a 128 GB RTX Spark. Nemotron 3.5 Super VL 120B-A12B: Unsloth's planning estimate for a modest context, including the OS, runtime, KV cache and vision file; not a measured run. This setup counts 32 GB of VRAM + 64 GB of RAM = 96 GB. Longer contexts and video need more. The full checker lets you change the route or recipe settings and lists the files and the published recipe.
Can an RTX 5090 run FLUX 3 Image Klein (int8, ComfyUI)?
A published route or recipe matches, with caveats, with 64 GB of system RAM in RigCheck's quick setup: Matches BFL's 8 GB guidance: the int8 files in ComfyUI; BFL: the int8 ComfyUI files run on 8 GB GPUs such as the RTX 3070. Image model, int8 text encoder and VAE together are 11.0 GB; by our reading, on a card smaller than that ComfyUI keeps them on the GPU one at a time and needs system RAM for the rest. BFL gives no RAM figure. The prompt upsampler BFL says not to skip is a 9B model (6.72 GB w4a8, 9.75 GB int8 or 18.8 GB BF16 in ComfyUI); BFL does not say whether its 8 GB figure includes it. When it does not fit with the image model, run it on its own first (BFL's README) or switch it off in the ComfyUI template, with weaker results. NVIDIA measured about 4 seconds per 1 MP image on an RTX 5090 (distilled, int8, 16 steps, ComfyUI, upsampler not included). The full checker lets you change the route or recipe settings and lists the files and the published recipe.
Other graphics cards
RTX 3090 · RTX 4090 · RTX 4060 Ti 16 GB · Radeon VII / Instinct MI50 16 GB · RTX 3060 12 GB · RTX 4070 Ti · Every 32 GB card · All graphics cards · Any other card or Mac
Sources
- NVIDIA GeForce RTX 5090 specifications: memory size
- DeepSeek V4.1 Flash (September release, 1017 update) in DwarfStar: the checker, its sources and assumptions
- FLUX 3 Image Klein (int8, ComfyUI): the checker, its sources and assumptions
- Gemma 4 (E2B, E4B, 12B, 26B-A4B, 31B) in llama.cpp: the checker, its sources and assumptions
- GLM-5.3-Flash (Unsloth GGUF): the checker, its sources and assumptions
- Kolibri-1 (official FP8 route): the checker, its sources and assumptions
- MiniMax H3 (video + audio): the checker, its sources and assumptions
- Mistral Large 4: the checker, its sources and assumptions
- Mistral Medium 3.5 in llama.cpp: the checker, its sources and assumptions
- Nemotron 3.5 (Bolt 74B-A9B, Super VL 120B-A12B): the checker, its sources and assumptions
- Qwen3.8-27B in llama.cpp: the checker, its sources and assumptions
- Qwen3.8-Flash-Next in llama.cpp: the checker, its sources and assumptions
- Qwen-Image-2.1 (image generation): the checker, its sources and assumptions
- Qwen3.8-Flash-Next in the Strata engine: the checker, its sources and assumptions