Which model for 2 × RTX 4090 · 48 Go?
48 GB usable for the model. 90 open-weight models fit in memory.
The margin goes to context: count ~1 GB per 8k tokens for a 7B model. A small margin means a short context.
Unlocked by this setup
These models fit on no smaller setup in the catalogue (the next one down offers 40 usable GB), even in their lightest quantization.
| Model | Elo | Quantization | RAM required | Margin | Memory headroom | Run |
|---|---|---|---|---|---|---|
| Llama 3.3 70BMeta · 70.6B | 1318 | Q4_K_M | 46 Go | +2 Go | tight | ollama run llama3.3:70b |
| Qwen2.5 72BAlibaba · 72.7B | 1303 | Q4_K_M | 48 Go | +0 Go | tight | ollama run qwen2.5:72b |
| Llama 3.1 Nemotron 70BNVIDIA · 70.6B | 1299 | Q4_K_M | 46 Go | +2 Go | tight | ollama run nemotron:70b |
Also fit on a smaller setup
| Model | Elo | Quantization | RAM required | Margin | Memory headroom | Run |
|---|---|---|---|---|---|---|
| Gemma 4 31BGoogle · 31B | 1452 | Q8_0 | 39 Go | +9 Go | tight | Hugging Face ↗ |
| Gemma 4 26B-A4BGoogle · 26B | 1438 | Q8_0 | 33 Go | +15 Go | comfortable | Hugging Face ↗ |
| Qwen3.8 27BAlibaba · 27B | 1438 | Q8_0 | 34 Go | +14 Go | tight | Hugging Face ↗ |
| Gemma 3 27BGoogle · 27B | 1365 | Q4_K_MQ8_0 also fits · 34 Go | 19 Go | +29 Go | comfortable | ollama run gemma3:27b |
| Mistral Small 3.2Mistral AI · 24B | 1356 | Q4_K_MQ8_0 also fits · 30 Go | 17 Go | +31 Go | comfortable | ollama run mistral-small3.2 |
| Qwen3 32BAlibaba · 32.8B | 1347 | Q4_K_MQ8_0 also fits · 41 Go | 23 Go | +25 Go | comfortable | ollama run qwen3:32b |
| Gemma 3 12BGoogle · 12B | 1342 | Q4_K_MFP16 also fits · 28 Go | 9 Go | +39 Go | comfortable | ollama run gemma3:12b |
| QwQ 32BAlibaba · 32.8B | 1336 | Q4_K_MQ8_0 also fits · 41 Go | 23 Go | +25 Go | comfortable | ollama run qwq |
| Gemma 3n E4BGoogle · 4B | 1318 | Q4_K_MFP16 also fits · 11 Go | 5 Go | +43 Go | comfortable | ollama run gemma3n:e4b |
| gpt-oss-20bOpenAI · 21B | 1317 | MXFP4 | 17 Go | +31 Go | comfortable | ollama run gpt-oss:20b |
| Nemotron 3 Nano 30B-A3BNVIDIA · 30B | 1314 | Q8_0 | 37 Go | +11 Go | tight | Hugging Face ↗ |
| Granite 4.1 8BIBM · 8B | 1304 | FP16 | 20 Go | +28 Go | comfortable | Hugging Face ↗ |
| Gemma 3 4BGoogle · 4B | 1303 | Q4_K_MFP16 also fits · 11 Go | 5 Go | +43 Go | comfortable | ollama run gemma3:4b |
| Qwen2.5-Coder 32BAlibaba · 32.8B | 1271 | Q4_K_MQ8_0 also fits · 41 Go | 23 Go | +25 Go | comfortable | ollama run qwen2.5-coder:32b |
| Aya Expanse 32BCohere · 32.3B | 1267 | Q4_K_MQ8_0 also fits · 40 Go | 22 Go | +26 Go | comfortable | ollama run aya-expanse:32b |
| Phi-4Microsoft · 14.7B | 1256 | Q4_K_MFP16 also fits · 34 Go | 11 Go | +37 Go | comfortable | ollama run phi4 |
| Phi-4 ReasoningMicrosoft · 14.7B | 1256 | Q4_K_MFP16 also fits · 34 Go | 11 Go | +37 Go | comfortable | ollama run phi4-reasoning |
| OLMo 2 32BAllen AI · 32B | 1252 | Q8_0 | 40 Go | +8 Go | tight | Hugging Face ↗ |
| CodestralMistral AI · 22.2B | 1240 | Q4_K_MQ8_0 also fits · 28 Go | 16 Go | +32 Go | comfortable | ollama run codestral |
| Ministral 8BMistral AI · 8.02B | 1238 | FP16 | 20 Go | +28 Go | comfortable | Hugging Face ↗ |
| Llama 3.1 8BMeta · 8.03B | 1211 | Q4_K_MFP16 also fits · 20 Go | 7 Go | +41 Go | comfortable | ollama run llama3.1:8b |
| Mixtral 8x7BMistral AI · 46.7B | 1197 | Q4_K_M | 31 Go | +17 Go | comfortable | ollama run mixtral:8x7b |
| Llama 3.2 3BMeta · 3.21B | 1167 | Q4_K_MFP16 also fits · 9 Go | 4 Go | +44 Go | comfortable | ollama run llama3.2:3b |
| Llama 3.2 1BMeta · 1.24B | 1111 | Q4_K_MFP16 also fits · 5 Go | 3 Go | +45 Go | comfortable | ollama run llama3.2:1b |
| Mistral 7BMistral AI · 7.25B | 1110 | Q4_K_MFP16 also fits · 18 Go | 7 Go | +41 Go | comfortable | ollama run mistral:7b |
Not yet ranked by the Arena
No Arena Elo linked: recent release, model absent from the Arena, or a match we have not made yet.
| Model | Elo | Quantization | RAM required | Margin | Memory headroom | Run |
|---|---|---|---|---|---|---|
| Qwen3 0.6BAlibaba · 0.6B | — | Q4_K_MFP16 also fits · 3 Go | 2 Go | +46 Go | comfortable | ollama run qwen3:0.6b |
| Gemma 3 1BGoogle · 1B | — | Q4_K_MFP16 also fits · 4 Go | 3 Go | +45 Go | comfortable | ollama run gemma3:1b |
| Phi-4 MiniMicrosoft · 3.8B | — | Q4_K_MFP16 also fits · 10 Go | 4 Go | +44 Go | comfortable | ollama run phi4-mini |
| Qwen3 4BAlibaba · 4.02B | — | Q4_K_MFP16 also fits · 11 Go | 5 Go | +43 Go | comfortable | ollama run qwen3:4b |
| Gemma 4 E2BGoogle · 2B | — | FP16 | 6 Go | +42 Go | comfortable | Hugging Face ↗ |
| CodeGemma 7BGoogle · 8.5B | — | Q4_K_MFP16 also fits · 21 Go | 7 Go | +41 Go | comfortable | ollama run codegemma |
| Command R7BCohere · 8B | — | Q4_K_MFP16 also fits · 20 Go | 7 Go | +41 Go | comfortable | ollama run command-r7b |
| DeepSeek-R1 Distill 8BDeepSeek · 8.03B | — | Q4_K_MFP16 also fits · 20 Go | 7 Go | +41 Go | comfortable | ollama run deepseek-r1:8b |
| Granite 3.3 8BIBM · 8.1B | — | Q4_K_MFP16 also fits · 20 Go | 7 Go | +41 Go | comfortable | ollama run granite3.3:8b |
| Hermes 3 8BNous Research · 8.03B | — | Q4_K_MFP16 also fits · 20 Go | 7 Go | +41 Go | comfortable | ollama run hermes3:8b |
| MiniCPM5-2BOpenBMB · 2.5B | — | FP16 | 7 Go | +41 Go | comfortable | Hugging Face ↗ |
| Qwen3 8BAlibaba · 8.2B | — | Q4_K_MFP16 also fits · 20 Go | 7 Go | +41 Go | comfortable | ollama run qwen3:8b |
| Falcon 3 10BTII · 10.3B | — | Q4_K_MFP16 also fits · 25 Go | 8 Go | +40 Go | comfortable | ollama run falcon3:10b |
| Granite 4.1 3BIBM · 3B | — | FP16 | 9 Go | +39 Go | comfortable | Hugging Face ↗ |
| Ministral 3 3BMistral AI · 3B | — | FP16 | 9 Go | +39 Go | comfortable | Hugging Face ↗ |
| SmolLM3 3BHugging Face · 3.1B | — | FP16 | 9 Go | +39 Go | comfortable | Hugging Face ↗ |
| Mistral Nemo 12BMistral AI · 12.2B | — | Q4_K_MFP16 also fits · 29 Go | 10 Go | +38 Go | comfortable | ollama run mistral-nemo |
| Gemma 4 E4BGoogle · 4B | — | FP16 | 11 Go | +37 Go | comfortable | Hugging Face ↗ |
| Qwen3 14BAlibaba · 14.8B | — | Q4_K_MFP16 also fits · 35 Go | 11 Go | +37 Go | comfortable | ollama run qwen3:14b |
| DeepSeek-Coder-V2 LiteDeepSeek · 15.7B | — | Q4_K_MFP16 also fits · 37 Go | 12 Go | +36 Go | comfortable | ollama run deepseek-coder-v2:16b |
| StarCoder 2BigCode · 16B | — | Q4_K_MFP16 also fits · 37 Go | 12 Go | +36 Go | comfortable | ollama run starcoder2:15b |
| Devstral SmallMistral AI · 24B | — | Q4_K_MQ8_0 also fits · 30 Go | 17 Go | +31 Go | comfortable | ollama run devstral |
| Lucie 7BOpenLLM France · 6.7B | — | FP16 | 17 Go | +31 Go | comfortable | Hugging Face ↗ |
| Magistral SmallMistral AI · 24B | — | Q4_K_MQ8_0 also fits · 30 Go | 17 Go | +31 Go | comfortable | ollama run magistral |
| OLMo 3 7BAllen AI · 7B | — | FP16 | 17 Go | +31 Go | comfortable | Hugging Face ↗ |
| Apertus 8BSwiss AI · 8B | — | FP16 | 20 Go | +28 Go | comfortable | Hugging Face ↗ |
| Ministral 3 8BMistral AI · 8B | — | FP16 | 20 Go | +28 Go | comfortable | Hugging Face ↗ |
| InternLM3 8BShanghai AI Lab · 8.8B | — | FP16 | 21 Go | +27 Go | comfortable | Hugging Face ↗ |
| EuroLLM 9BUTTER Project · 9.15B | — | FP16 | 22 Go | +26 Go | comfortable | Hugging Face ↗ |
| DeepSeek-R1 Distill 32BDeepSeek · 32.8B | — | Q4_K_MQ8_0 also fits · 41 Go | 23 Go | +25 Go | comfortable | ollama run deepseek-r1:32b |
| Gemma 4 12BGoogle · 12B | — | FP16 | 28 Go | +20 Go | comfortable | Hugging Face ↗ |
| Ministral 3 14BMistral AI · 14B | — | FP16 | 33 Go | +15 Go | comfortable | Hugging Face ↗ |
| Qwen3.6 27BAlibaba · 27B | — | Q8_0 | 34 Go | +14 Go | tight | Hugging Face ↗ |
| Qwen3.6 35B-A3BAlibaba · 35B | — | Q8_0 | 43 Go | +5 Go | tight | Hugging Face ↗ |
Image, video, audio and 3D
These models do not generate text, so the text Arena does not rank them. Indicative order, not a quality ranking.
| Model | Elo | Quantization | RAM required | Margin | Memory headroom | Run |
|---|---|---|---|---|---|---|
| Kokoro-82MHexgrad · 0.08B | — | FP16 | 2 Go | +46 Go | comfortable | Hugging Face ↗ |
| ChatterboxResemble AI · 0.5B | — | FP16 | 3 Go | +45 Go | comfortable | Hugging Face ↗ |
| Parakeet TDT 0.6B v3NVIDIA · 0.6B | — | FP16 | 3 Go | +45 Go | comfortable | Hugging Face ↗ |
| TripoSRStability AI · 0.5B | — | FP16 | 3 Go | +45 Go | comfortable | Hugging Face ↗ |
| Stable Fast 3DStability AI · 1B | — | FP16 | 4 Go | +44 Go | comfortable | Hugging Face ↗ |
| TRELLISMicrosoft · 1.2B | — | FP16 | 5 Go | +43 Go | comfortable | Hugging Face ↗ |
| Whisper large-v3OpenAI · 1.55B | — | FP16 | 5 Go | +43 Go | comfortable | Hugging Face ↗ |
| Dia 1.6BNari Labs · 1.6B | — | FP16 | 6 Go | +42 Go | comfortable | Hugging Face ↗ |
| Hunyuan3D 2.1Tencent · 3B | — | FP16 | 9 Go | +39 Go | comfortable | Hugging Face ↗ |
| Orpheus 3BCanopy Labs · 3B | — | FP16 | 9 Go | +39 Go | comfortable | Hugging Face ↗ |
| Voxtral Mini 3BMistral AI · 3B | — | FP16 | 9 Go | +39 Go | comfortable | Hugging Face ↗ |
| YuE2-3Bm-a-p (Multimodal Art Projection) · 3B | — | FP16 | 9 Go | +39 Go | comfortable | Hugging Face ↗ |
| FLUX.2 [klein] 4BBlack Forest Labs · 4B | — | FP16 | 11 Go | +37 Go | comfortable | Hugging Face ↗ |
| TRELLIS.2Microsoft · 4B | — | FP16 | 11 Go | +37 Go | comfortable | Hugging Face ↗ |
| CogVideoX 5BZhipu AI · 5B | — | FP16 | 13 Go | +35 Go | comfortable | Hugging Face ↗ |
| Wan 2.2 5BAlibaba · 5B | — | FP16 | 13 Go | +35 Go | comfortable | Hugging Face ↗ |
| Ming-Image 0.1 DesignAnt Group · 6B | — | FP16 | 15 Go | +33 Go | comfortable | Hugging Face ↗ |
| Z-Image TurboAlibaba · 6B | — | FP16 | 15 Go | +33 Go | comfortable | Hugging Face ↗ |
| HunyuanVideo 1.5Tencent · 8.3B | — | FP16 | 20 Go | +28 Go | comfortable | Hugging Face ↗ |
| GLM-ImageZhipu AI · 9B | — | FP16 | 22 Go | +26 Go | comfortable | Hugging Face ↗ |
| Mochi 1Genmo · 10B | — | FP16 | 24 Go | +24 Go | comfortable | Hugging Face ↗ |
| FLUX.1 [dev]Black Forest Labs · 12B | — | FP16 | 28 Go | +20 Go | comfortable | Hugging Face ↗ |
| LTX-2.3Lightricks · 22B | — | Q8_0 | 28 Go | +20 Go | comfortable | Hugging Face ↗ |
| Voxtral Small 24BMistral AI · 24B | — | Q8_0 | 30 Go | +18 Go | comfortable | Hugging Face ↗ |
| Wan 2.2 A14BAlibaba · 27B | — | Q8_0 | 34 Go | +14 Go | tight | Hugging Face ↗ |
| HiDream-I1HiDream · 17B | — | FP16 | 39 Go | +9 Go | tight | Hugging Face ↗ |
| FLUX.2 [dev]Black Forest Labs · 32B | — | Q8_0 | 40 Go | +8 Go | tight | Hugging Face ↗ |
| Qwen-Image 2512Alibaba · 20B | — | FP16 | 46 Go | +2 Go | tight | Hugging Face ↗ |
An “ollama run” command without a suffix pulls the Q4_K_M build from the Ollama library: the row then shows that quantization and its memory, and flags whether FP16 or Q8_0 would also fit. Exception: a model published in a native format (MXFP4 for gpt-oss, FP8 for DeepSeek V4.1) exists in one build only, which is what the command loads. via Hugging Face or an explicit Ollama tag.
Share this verdict
One link, the same shortlist, the same reasons.