Qwen3.8 27B
Local OKExcellentA · 43
Best for: Coding, Agent, Vision, Long context·Pop: 95/100
Perf: est. 89 tok/s, typical error 5% · first token ~0.3s
Fits in 32 GB VRAM with room to spare. Best for coding, agent, vision, long context on RTX 5090.
Best Local AI Models for RTX 5090 (32GB)
The RTX 5090 pairs 32GB of GDDR7 with 1,792 GB/s of bandwidth, the most of any consumer card. It runs 8B models at est. 213 tok/s, typical error 5% and loads 70B parameter models at Q3 quantization. No other consumer GPU holds a 70B model locally.
Quick answer
The best local LLM for the RTX 5090 is Qwen3.8 27B (est. 89 tok/s, typical error 5%) on its 32GB VRAM. It uses ~16.5GB of VRAM; the RTX 5090 handles up to 32B parameter models at Q4. A 32B model at Q4 runs at est. 69 tok/s, typical error 5%. (ModelFit, 2026)
Sizing rule: a Q4 model needs about 0.6 GB of VRAM per billion parameters, and ModelFit budgets 90% of this card's 32GB for weights, context, and KV-cache. The per-size table below uses that same budget. Other strong fits: Qwen3.6 35B-A3B (35B, ~22GB) and Gemma 4 26B-A4B (26B, ~16GB). On this card Qwen3.8 27B runs at est. 89 tok/s, typical error 5%. What it will not run: a 70B model does not fit in VRAM alone, so it spills to system RAM over PCIe and slows sharply.
$ollama run qwen3.8:27b
- Top pick
- Qwen3.8 27B
- Decode speed
- est. 89 tok/s, typical error 5%
- VRAM needed
- ~16.5 GB
Speeds are ModelFit estimates from memory bandwidth and model size, not measured benchmarks. Typical error: the median gap between the ModelFit speed model and measured runs on hardware the model was not fitted on, for this backend and model size.
Cite this page: ModelFit, RTX 5090 32GB Local LLM: 32B at est. 69 tok/s, 70B at Q3, https://modelfit.io/gpu/rtx-5090/, updated October 2026, CC BY 4.0.
Specs
VRAM
32 GB GDDR7
Speed (8B Q4)
est. 213 tok/stypical error 5%
Bandwidth
1792 GB/s
Architecture
Blackwell
Max model size
Up to 70B parameter models
Compatibility
10 excellent, 0 workable
Vendor
Table 1
RTX 5090 estimated tokens/sec by model size
Q4_K_M · ModelFit estimate| Model size | Est. speed | Fit on 32GB |
|---|---|---|
| 7B | est. 237 tok/stypical error 5% | Fits in VRAM |
| 14B | est. 141 tok/stypical error 15% | Fits in VRAM |
| 20B MoE (3.6B active) | est. 266 tok/stypical error 15% | Fits in VRAM |
| 32B | est. 69 tok/stypical error 5% | Fits in VRAM |
| 35B MoE (3B active) | est. 298 tok/stypical error 15% | Fits in VRAM |
| 70B | est. 3 tok/stypical error 10% | CPU offload (slow) |
| 120B MoE (5.1B active) | est. 74 tok/stypical error 15% | CPU offload (slow) |
Fig. 1RTX 5090: estimated speed by model size
Decode speed, generic Q4_K_M model of each size at 4K context, 1792 GB/s, typical error 5-15%
Fits in 32 GBCPU offload (slow)
7B237
14B141
20B MoE (3.6B active)266
32B69
35B MoE (3B active)298
70B3
120B MoE (5.1B active)74
tok/s est.ModelFit estimates, not measured benchmarks: decode speed of a generic Q4_K_M model of each size at 4K context on the RTX 5090's 1792 GB/s bandwidth. MoE rows scale by active parameters (decode reads only the active experts), so a 35B MoE runs far faster than a dense 32B. "CPU offload" sizes exceed the 32GB VRAM; dense models slow to a crawl there, MoE models degrade less because hot experts stay GPU-resident. What typical error means.
Fig. 2
Context costs VRAM too.
Qwen3.8 27B loads ~16.5 GB of weights; at 16k context the KV cache adds ~1.0 GB (still fits the ~29 GB usable VRAM), and at 64k it adds ~4.0 GB (still fits).
Weights 17 GBKV cache · 16k context 1.0 GBHeadroom 11 GBRuntime reserve 3.2 GB
The model leaves about 11 GB of the usable VRAM budget free after weights and 16k context.
KV-cache figures assume an fp16 cache, the llama.cpp/Ollama default. Standard GQA models use a size-class estimate (8 KV heads x 128 head dim class); hybrid linear-attention models (Qwen3.5/3.6, Qwen3-Next) use the exact per-token cost from their published config, since only their sparse full-attention layers cache KV. A q8_0 KV cache roughly halves either figure. Estimates, not measurements.
Get told when a model beats Qwen3.8 27B on your RTX 5090
One email when a new open-weight model is a better fit for 32 GB of VRAM, plus the Thursday weekly.
You'll also get the weekly ModelFit email. By signing up you agree to our Privacy Policy. Unsubscribe from either email anytime, in one click.
Where to buy the RTX 5090
Storage & accessories for your model library
- Internal NVMe SSD · 2TBA Gen4 M.2 drive keeps your whole GGUF and quant collection on fast local storage, loading models straight off NVMe.Check price on Amazon
- USB4 NVMe Enclosure40Gbps external storage fast enough to run models from. Pair it with an M.2 drive for a portable model vault.Check price on Amazon
ModelFit may earn a commission on purchases through these links, at no extra cost to you.
RTX 5090 VRAM for AI: what actually fits?
32GB GDDR7 at 1,792 GB/s is the highest bandwidth of any consumer card. This is enough to load 32B models at Q6 or even Q8 quantization with room to spare, delivering higher quality than Q4 on smaller cards. The 5090 also handles multi-model setups: run a 14B chat model alongside a 7B coding model simultaneously. For the largest 70B models at Q3_K_M (~30GB), it fits with about 1GB headroom. No other consumer card runs a 70B model at this quality.
Table 2
RTX 5090 vs similar GPUs
Recommended models
registry-verified10 modelsQwen3.6 35B-A3B
Local OKOKB · 24
Best for: Reasoning, Coding, Agents·Pop: 88/100
Perf: est. 346 tok/s, typical error 15% · first token ~0.1s
Fits in 32 GB VRAM with room to spare. Best for reasoning, coding, agents on RTX 5090.
Gemma 4 26B-A4B
Local OKExcellentA · 44
Best for: Chat, Coding, Multimodal·Pop: 86/100
Perf: est. 260 tok/s, typical error 15% · first token ~0.1s
Fits in 32 GB VRAM with room to spare. Best for chat, coding, multimodal on RTX 5090.
Gemma 4 31B
Local OKExcellentA · 31
Best for: Quality, Coding, Multimodal·Pop: 84/100
Perf: est. 73 tok/s, typical error 5% · first token ~0.3s
Fits in 32 GB VRAM with room to spare. Best for quality, coding, multimodal on RTX 5090.
Qwen3.8 27B (Q8)
Local OKOKB · 6
Best for: Coding, Agent, Vision, Long context·Pop: 95/100
Perf: est. 55 tok/s, typical error 5% · first token ~1.4s
Fits in 32 GB VRAM with room to spare. Best for coding, agent, vision, long context on RTX 5090.
Ornith 1.0 35B
Local OKOKB · 26
Best for: Agentic coding·Pop: 72/100
Perf: est. 350 tok/s, typical error 15% · first token ~0.1s
Fits in 32 GB VRAM with room to spare. Best for agentic coding on RTX 5090.
Flint 40B-A2.5B
Local OKOKB · 20
Best for: Agentic coding, Long context·Pop: 72/100
Perf: est. 363 tok/s, typical error 15% · first token ~0.1s
Fits in 32 GB VRAM with room to spare. Best for agentic coding, long context on RTX 5090.
Laguna XS 2.1
Local OKOKB · 30
Best for: Agentic coding, Long-horizon tasks·Pop: 72/100
Perf: est. 332 tok/s, typical error 15% · first token ~0.1s
Fits in 32 GB VRAM with room to spare. Best for agentic coding, long-horizon tasks on RTX 5090.
Nemotron 3.5 Lightning 30B-A3B (IQ4_XS)
Local OKExcellentA · 34
Best for: Agentic, Coding, Long context·Pop: 78/100
Perf: est. 300 tok/s, typical error 15% · first token ~0.1s
Fits in 32 GB VRAM with room to spare. Best for agentic, coding, long context on RTX 5090.
Nemotron 3.5 Lightning 30B-A3B
Local OKOKB · 12
Best for: Agentic, Coding, Long context·Pop: 78/100
Perf: est. 250 tok/s, typical error 15% · first token ~0.1s
Fits in 32 GB VRAM with room to spare. Best for agentic, coding, long context on RTX 5090.
Video and image models on the RTX 5090
Runs · Stated: the linked source says a GPU with that much VRAM runs the model. Runs · Published: a dated run on the named card with that much VRAM. unknown: no source yet. Every VRAM tier: video · image.
Video
| Model | Fits | Quality | App |
|---|---|---|---|
| HunyuanVideo-1.5Tencent Hunyuan Community License | Runs · Stated24 GB VRAM · ComfyUI native workflow · ComfyUI docs: HunyuanVideo 1.5 | #38 of 48 in text to video4,292 votessnapshot 2026-09-22 | ComfyUI |
| LTX-2.3LTX-2 Community License Agreement | Runs · Stated24 GB VRAM · two-stage production at 720p to 1080p · LTX blog: hardware for AI video models | No independent score yet | Lightricks/LTX-2 ltx-pipelines, ComfyUI |
| LTX-2.5LTX-2.x Community License Agreement | Runs · Stated32 GB VRAM · official requirement, plus 32 GB system RAM · LTX-2.5 page (ltx.io) | No independent score yet | Lightricks/LTX-2 ltx-pipelines, ComfyUI |
| MiniMax H3MiniMax H3 Community License Agreement | Runs · Stated24 GB VRAM · Diffusers int8 with streamed blocks, about 75 GB system RAM · Diffusers docs: MiniMax-H3 memory | #8 of 48 in text to video11,417 votessnapshot 2026-09-22 | stable-diffusion.cpp, ComfyUI |
| Mochi 1 previewApache License 2.0 | Runs · Stated22 GB VRAM · Diffusers bf16 variant with CPU offload · Mochi 1 model cardComplete: Official Diffusers BF16 variant40.0 GB download | #48 of 48 in text to video5,893 votessnapshot 2026-09-22 | Diffusers, ComfyUI |
| Wan2.1Apache License 2.0 | Runs · Stated13 GB VRAM · T2V 14B in Diffusers, bf16 with group offloading to system RAM · Diffusers docs: Wan pipeline | No independent score yet | ComfyUI |
| Wan2.2 A14BApache License 2.0 | Runs · PublishedRTX 5060 Ti · 16 GB VRAM · ComfyUI fp16 I2V with 4-step LoRAs, 64 GB system RAM · ComfyUI issue #10662 | #42 of 48 in text to video10,458 votessnapshot 2026-09-22 | ComfyUI |
| Wan2.2 TI2V 5BApache License 2.0 | Runs · Stated24 GB VRAM · official generate.py at 1280x704 with offload flags · Wan2.2-TI2V-5B model card | No independent score yet | ComfyUI |
Quality: blind human votes from Arena (arena.ai), CC BY 4.0, snapshot 2026-09-22. Scores are for the Arena checkpoint, not your local quantized build.
Image
| Model | Fits | Quality | App |
|---|---|---|---|
| FLUX 3 Image KleinFLUX Kommunity License v1.0 | Runs · Stated8 GB VRAM · official ComfyUI int8 files · FLUX 3 Image Klein README (Black Forest Labs) | No independent score yet | Diffusers |
| FLUX.1 [dev]FLUX.1 [dev] Non-Commercial License v1.1.1 | Runs · Stated10 GB VRAM · FP4 on RTX 50 (Blackwell) · Black Forest Labs blog: FLUX with NVIDIA | #78 of 82 in text to image49,221 votessnapshot 2026-10-07 | Diffusers, ComfyUI |
| FLUX.2 [dev]FLUX Non-Commercial License | Runs · Stated24 GB VRAM · 4-bit quantization in Diffusers · FLUX.2 dev Diffusers guide (Black Forest Labs) | #38 of 82 in text to image77,988 votessnapshot 2026-10-07 | Diffusers, ComfyUI, stable-diffusion.cpp |
| FLUX.2 [klein]Apache 2.0 (4B), FLUX Non-Commercial License (9B) | Runs · Stated29 GB VRAM · Klein 9B · FLUX.2 [klein] 9B model cardComplete: Klein 9B BF1634.7 GB download | No independent score yet | Diffusers, stable-diffusion.cpp |
| Krea 2 TurboKrea 2 Community License Agreement v.1 | Runs · PublishedRTX 4070 · 12 GB VRAM · ComfyUI fp8, dynamic VRAM paging, 8 steps · ComfyUI issue #15365 | No independent score yet | Diffusers, stable-diffusion.cpp |
| Qwen-Image-2.1Qwen Research License Agreement | Runs · PublishedRTX 5080 · 16 GB VRAM · stable-diffusion.cpp Q8_0, 1024x1024 · stable-diffusion.cpp issue #2041 | #19 of 82 in text to image11,441 votessnapshot 2026-10-07 | Diffusers, stable-diffusion.cpp |
| SDXL Base 1.0CreativeML Open RAIL++-M License | Runs · Stated8 GB VRAM · SDXL 1.0, no app named · SDXL 1.0 announcement (Stability AI) | No independent score yet | ComfyUI, stable-diffusion.cpp |
| Stable Diffusion 3.5 LargeStability AI Community License | Runs · Stated11 GB VRAM · FP8 TensorRT build on RTX 40 and 50 · Stability AI: SD3.5 TensorRT models | #81 of 82 in text to image23,396 votessnapshot 2026-10-07 | ComfyUI |
| Z-Image-TurboApache 2.0 | Runs · Stated16 GB VRAM · stated for consumer GPUs · Z-Image-Turbo model card | #59 of 82 in text to image24,808 votessnapshot 2026-10-07 | Diffusers, stable-diffusion.cpp |
Quality: blind human votes from Arena (arena.ai), CC BY 4.0, snapshot 2026-10-07. Scores are for the Arena checkpoint, not your local quantized build.
Can I run it on the RTX 5090?
Per-model verdict pages for this exact card: weights, KV cache by context length, and estimated tok/s, each derived from the same engine the recommendations above use.
- Qwen3 8B Q4_K_Mfits
- Qwen3 14B Q4_K_Mfits
- GPT-OSS 20B MXFP4fits
- Qwen3.8 27B Q4_K_Mfits
- Llama 3.3 70B Instruct Q4_K_Mtight
- Qwen3.5 35B-A3B Instruct Q4_K_Mfits
- Gemma 4 26B-A4B Q4_K_Mfits
- Qwen3.5 4B Instruct Q4_K_Mfits
- Gemma 4 E4B Q4_K_Mfits
- Gemma 3 4B Instruct Q4_K_Mfits
- Llama 3.1 8B Instruct Q4_K_Mfits
- Qwen3.5 9B Instruct Q4_K_Mfits
- Gemma 4 12B Q4_K_Mfits
- Mistral Nemo 12B Q4_K_Mfits
- Qwen3.5 9B Instruct (Q8) Q8_0fits
- Gemma 4 12B (Q8) Q8_0fits
- LFM2 24B-A2B Instruct Q4_K_Mfits
- Qwen3.5 27B Instruct Q4_K_Mfits
- Qwen3.6 27B Q4_K_Mfits
- Gemma 4 31B Q4_K_Mfits
- Qwen3.6 35B-A3B Q4_K_Mfits
- Qwen3.8 27B (Q8) Q8_0fits
Similar GPUs for local AI
Speeds are estimates for Qwen3 8B Q4_K_M, typical error 5%.
Compatible model families
- QwenAlibaba Cloud: Qwen3.8 27B local flagship (Apache 2.0, vision)
- LlamaMeta: Most popular open-weight model family
- DeepSeekDeepSeek AI: V4.1 Flash: multimodal 552B MoE, API only
- MistralMistral AI: Excellent performance-per-parameter ratio
- GemmaGoogle DeepMind: Gemma 4 E4B: 4.5B effective, runs on 8GB Macs
- PhiMicrosoft: Best quality-per-gigabyte at small sizes
RTX 5090 FAQ: common questions
How much VRAM does the RTX 5090 have for LLMs?
The RTX 5090 has 32GB GDDR7 VRAM with 1,792 GB/s bandwidth, the most of any consumer GPU. About 31GB is usable for models. This fits 70B models at Q3 quantization and 32B models at Q8 for maximum quality.
What size LLM can I run on an RTX 5090?
Up to 70B parameter models at Q3 quantization, or 32B models at Q8 for best quality. The 5090 is the only consumer GPU that can run Llama 3.1 70B locally. For 7B-8B models it delivers est. 213 tok/s, typical error 5% (14B lands around est. 141 tok/s, typical error 15%).
Is the RTX 5090 good for local AI in 2026?
The RTX 5090 has the most VRAM (32GB) and the highest bandwidth (1,792 GB/s) of any consumer GPU. At est. 213 tok/s, typical error 5% on 8B models, responses feel instant. Its 32GB VRAM unlocks 70B models that no other consumer card can handle. The 2026 memory shortage has pushed it far above its $1,999 launch MSRP, so it is only justified if you genuinely need large models.
RTX 5090 vs RTX 4090 for running AI models?
The RTX 5090 is 60% faster on 8B models (est. 213 tok/s, typical error 5% vs est. 133 tok/s) with 33% more VRAM (32 vs 24GB). It runs 70B models that the 4090 cannot. The 4090 is discontinued and both now trade well above MSRP, but the 5090 is the clear choice for new purchases in 2026.
Can the RTX 5090 replace cloud AI services?
For many use cases, yes. A 32B model on the 5090 rivals GPT-4 quality for coding and reasoning tasks, running at est. 69 tok/s, typical error 5% at Q4 with low latency and complete privacy. The main gap is context length: cloud models support 128K+ tokens while local 70B models are limited to 8K-16K.
How fast is a 27B-class model like Qwen3.8 27B on the RTX 5090?
ModelFit estimates a 32B model on the RTX 5090 at Q4_K_M runs at est. 69 tok/s, typical error 5%. The current 27B-class pick in the catalog is Qwen3.8 27B (ollama run qwen3.8:27b).
Related guides & benchmarks
More local AI guides
- How Much RAM (or VRAM) Do You Need for a Local LLM?The model-size-to-memory matrix: what each VRAM and RAM tier actually runs.
- Best LLM for MacBook (Apple Silicon)Unified-memory Macs run bigger models per dollar. Picks by RAM tier, M1 to M5.
- Local AI image generator on MacCompare image models by RAM and app.
- Local image generation on MacImage models and apps by Mac RAM.
Browse all NVIDIA GPUs for AI
Real-world calibration
Have this hardware? Share a measured run.
Measure a fixed Ollama reference model on your machine. This separate, opt-in contribution does not change ModelFit's third-party SWE-Bench leaderboard; results appear on the measured-speed leaderboard after the next site update.
$ npx -y @wecko-ai/modelfit bench --submit --cleanup
Want personalized recommendations?
Use our interactive wizard to compare models across Apple Silicon and NVIDIA GPUs.