Qwen3.8 27B
Local OKOKB · 24
Best for: Coding, Agent, Vision, Long context·Pop: 95/100
Perf: est. 51 tok/s, typical error 5% · first token ~0.3s
Fits in 24 GB VRAM with room to spare. Best for coding, agent, vision, long context on RTX 4090.
Best Local AI Models for RTX 4090 (24GB)
The RTX 4090 pairs 24GB of GDDR6X with 1,008 GB/s of memory bandwidth. It runs 8B chat models at est. 133 tok/s, typical error 5% and 32B reasoning models at est. 39 tok/s. Only the RTX 5090 is faster among consumer cards.
Quick answer
The best local LLM for the RTX 4090 is Qwen3.8 27B (est. 51 tok/s, typical error 5%) on its 24GB VRAM. It uses ~16.5GB of VRAM; the RTX 4090 handles up to 32B parameter models at Q4. A 32B model at Q4 runs at est. 39 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 24GB for weights, context, and KV-cache. The per-size table below uses that same budget. Other strong fits: Gemma 4 26B-A4B (26B, ~16GB) and Gemma 4 12B (Q8) (12B, ~12.8GB). On this card Qwen3.8 27B runs at est. 51 tok/s, typical error 5%. What it will not run: a 35B MoE (3B active) 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. 51 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 4090 24GB Local LLM (2026): Runs 32B at est. 39 tok/s, https://modelfit.io/gpu/rtx-4090/, updated October 2026, CC BY 4.0.
Specs
VRAM
24 GB GDDR6X
Speed (8B Q4)
est. 133 tok/stypical error 5%
Bandwidth
1008 GB/s
Architecture
Ada Lovelace
Max model size
Up to 32B parameter models
Compatibility
10 excellent, 0 workable
Vendor
Table 1
RTX 4090 estimated tokens/sec by model size
Q4_K_M · ModelFit estimate| Model size | Est. speed | Fit on 24GB |
|---|---|---|
| 7B | est. 150 tok/stypical error 5% | Fits in VRAM |
| 14B | est. 84 tok/stypical error 15% | Fits in VRAM |
| 20B MoE (3.6B active) | est. 172 tok/stypical error 15% | Fits in VRAM |
| 32B | est. 39 tok/stypical error 5% | Fits in VRAM |
| 35B MoE (3B active) | est. 148 tok/stypical error 15% | CPU offload (slow) |
| 70B | est. 2 tok/stypical error 10% | CPU offload (slow) |
| 120B MoE (5.1B active) | est. 46 tok/stypical error 15% | CPU offload (slow) |
Fig. 1RTX 4090: estimated speed by model size
Decode speed, generic Q4_K_M model of each size at 4K context, 1008 GB/s, typical error 5-15%
Fits in 24 GBCPU offload (slow)
7B150
14B84
20B MoE (3.6B active)172
32B39
35B MoE (3B active)148
70B2
120B MoE (5.1B active)46
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 4090's 1008 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 24GB 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 ~22 GB usable VRAM), and at 64k it adds ~4.0 GB (still fits).
Weights 17 GBKV cache · 16k context 1.0 GBHeadroom 4.1 GBRuntime reserve 2.4 GB
The model leaves about 4.1 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 4090
One email when a new open-weight model is a better fit for 24 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 4090
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 4090 VRAM for AI: what actually fits?
24GB GDDR6X at 1,008 GB/s gives the RTX 4090 real headroom. 32B models at Q4 (~20GB) load fully with 3GB left for KV cache. 14B models at Q5 or Q6 fit easily for higher quality inference. At est. 133 tok/s, typical error 5% with 8B models, the 4090 delivers near-instant responses. The only consumer card faster is the RTX 5090 (est. 213 tok/s, 32GB). Among 24GB cards, its 1,008 GB/s bandwidth is the highest ModelFit tracks.
Table 2
RTX 4090 vs similar GPUs
Recommended models
registry-verified10 modelsGemma 4 26B-A4B
Local OKOKB · 26
Best for: Chat, Coding, Multimodal·Pop: 86/100
Perf: est. 168 tok/s, typical error 15% · first token ~0.1s
Fits in 24 GB VRAM with room to spare. Best for chat, coding, multimodal on RTX 4090.
Gemma 4 12B (Q8)
Local OKExcellentA · 41
Best for: Chat, Coding, Multimodal·Pop: 80/100
Perf: est. 61 tok/s, typical error 15% · first token ~0.9s
Fits in 24 GB VRAM with room to spare. Best for chat, coding, multimodal on RTX 4090.
Qwen3.5 35B-A3B Instruct
Local OKOKB · 7
Best for: Reasoning, Coding, Agent scenarios·Pop: 90/100
Perf: est. 242 tok/s, typical error 15% · first token ~0.1s
Fits in 24 GB VRAM with room to spare. Best for reasoning, coding, agent scenarios on RTX 4090.
Nemotron 3.5 Lightning 30B-A3B (IQ4_XS)
Local OKOKB · 13
Best for: Agentic, Coding, Long context·Pop: 78/100
Perf: est. 200 tok/s, typical error 15% · first token ~0.1s
Fits in 24 GB VRAM with room to spare. Best for agentic, coding, long context on RTX 4090.
LFM2 24B-A2B Instruct
Local OKExcellentA · 35
Best for: Local AI agents, privacy-first tool calling, MCP workflows·Pop: 80/100
Perf: est. 257 tok/s, typical error 15% · first token ~0.1s
Fits in 24 GB VRAM with room to spare. Best for local ai agents, privacy-first tool calling, mcp workflows on RTX 4090.
Gemma 4 31B
Local OKOKB · 7
Best for: Quality, Coding, Multimodal·Pop: 84/100
Perf: est. 41 tok/s, typical error 5% · first token ~0.4s
Fits in 24 GB VRAM with room to spare. Best for quality, coding, multimodal on RTX 4090.
AREX-2 27B
Local OKOKB · 20
Best for: Agents, Deep research, Coding·Pop: 60/100
Perf: est. 50 tok/s, typical error 5% · first token ~0.3s
Fits in 24 GB VRAM with room to spare. Best for agents, deep research, coding on RTX 4090.
GLM-4.7-Flash 30B-A3B
Local OKOKB · 12
Best for: Coding, Agentic tasks, Reasoning·Pop: 78/100
Perf: est. 178 tok/s, typical error 15% · first token ~0.1s
Fits in 24 GB VRAM with room to spare. Best for coding, agentic tasks, reasoning on RTX 4090.
Laguna XS 2.1
Local OKOKB · 6
Best for: Agentic coding, Long-horizon tasks·Pop: 72/100
Perf: est. 226 tok/s, typical error 15% · first token ~0.1s
Fits in 24 GB VRAM with room to spare. Best for agentic coding, long-horizon tasks on RTX 4090.
Video and image models on the RTX 4090
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 · Stated24 GB VRAM · LTX Desktop app, 32 GB system RAM · LTX Desktop README (Lightricks) | 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 · Stated6 GB VRAM · stable-diffusion.cpp with GGUF weights · stable-diffusion.cpp FLUX docs | #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 · Stated13 GB VRAM · Klein 4B · FLUX.2 [klein] 4B model cardComplete: Klein 4B BF1616.0 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 4090 · 24 GB VRAM · Diffusers bf16 with model CPU offload, 1024x1024, 50 GB free system RAM required · MindrLabs Qwen-Image-2.1 README | #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 4090?
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_Mno
- 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_Mtight
- Qwen3.8 27B (Q8) Q8_0tight
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 4090 FAQ: common questions
How much VRAM does the RTX 4090 have for LLMs?
The RTX 4090 has 24GB GDDR6X VRAM with 1,008 GB/s bandwidth. About 23GB is usable for models. It runs 32B models at Q4 with room for 8K+ context windows.
What size LLM can I run on an RTX 4090?
Up to 32B parameter models at Q4 quantization. This includes DeepSeek-R1 32B, Qwen 2.5 32B, and larger reasoning models. For 70B models, you need Q2 or dual GPUs, or step up to the RTX 5090.
Is the RTX 4090 worth its 2026 price for AI?
Rarely. The card is discontinued and its street price has climbed well past launch MSRP, while it is only 14% faster than a used RTX 3090, which has the same 24GB VRAM for a fraction of the money. The 4090 is worth it if you also game at 4K or need the absolute fastest 24GB card. For pure AI value, the 3090 wins.
RTX 4090 vs RTX 5090 for AI: which is better?
The RTX 5090 is 60% faster on 8B models (est. 213 tok/s, typical error 5% vs est. 133 tok/s) with 8GB more VRAM (32 vs 24GB), enabling 70B models. The 5090 also costs more in 2026 now that the 4090 is discontinued and only scalped stock remains. If buying new, the 5090 is the clear choice.
Can the RTX 4090 run 70B models?
Not at full quality. A 70B Q4 model needs ~42GB VRAM. The 4090 has 24GB, so you would need Q2 quantization (lower quality) or run with partial CPU offloading (much slower). For 70B, the RTX 5090 (32GB) is recommended.
How fast is a 27B-class model like Qwen3.8 27B on the RTX 4090?
ModelFit estimates a 32B model on the RTX 4090 at Q4_K_M runs at est. 39 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.
Time a model that fits your machine with Ollama, using the same prompts as every other run. The CLI shows your result and asks before it sends anything; your machine then joins the speed test leaderboard. It does not change the quality benchmark.
$ npx @modelfit/cli@latest bench
Want personalized recommendations?
Use our interactive wizard to compare models across Apple Silicon and NVIDIA GPUs.