Best LLMs for 32 GB of VRAM
| Text models: 233 | Visual models: 15 | Sized for RTX 5090 | Strongest first |
Text models for 32 GB
| Model | Best fit | Context | Speed | Strongest at |
|---|---|---|---|---|
| Qwen3.8 27B | 8-bit | 24K | 41 tok/s | Chat, Coding, Agents, Docs |
| Muse Glimmer 30B | 6-bit | 128K | 50 tok/s | Docs |
| Qwen3.6 35B A3B | 6-bit | 96K | 227 tok/s | Chat, Coding, Agents, Docs |
| Gemma 4 31B | 6-bit | 112K | 45 tok/s | Chat, Coding, Agents, Docs |
| Gemma 4 26B A4B | 8-bit | 256K | 199 tok/s | Chat, Coding, Agents, Docs |
| Qwen3.5 9B | 8-bit | 256K | 120 tok/s | Chat, Coding, Agents, Docs |
| GLM-4.7 Flash | 6-bit | 104K | 209 tok/s | Chat, Coding, Agents |
| Gemma 4 12B | 8-bit | 256K | 86 tok/s | Chat, Coding, Docs |
| Nemotron 3.5 Lightning 30B A3B | 6-bit | 256K | 195 tok/s | Agents, Docs |
| Devstral Small 2 24B | FP8 | 32K | 47 tok/s | Coding |
| LFM2.5 24B A2B | 8-bit | 128K | 244 tok/s | Agents |
| Mistral Small 3.2 24B | 8-bit | 36K | 48 tok/s | Chat, Coding, Agents |
| Ministral 3 14B | 8-bit | 96K | 80 tok/s | Coding, Agents, Docs |
| Gemma 4 E4B | 8-bit | 128K | 177 tok/s | Chat, Coding, Docs |
| gpt-oss 20B | MXFP4 | 128K | 287 tok/s | Agents |
| Ministral 3 8B | 8-bit | 152K | 123 tok/s | Docs |
| Olmo 3.1 32B | 6-bit | 64K | 44 tok/s | Chat, Coding, Agents, Docs |
| Llama 3.1 8B | 8-bit | 128K | 137 tok/s | Chat, Coding, Agents, Docs |
| Qwen3 14B | 8-bit | 40K | 77 tok/s | Not rated |
| Qwen3 VL 8B Instruct | 8-bit | 144K | 125 tok/s | Text and images |
| Qwen3 8B | 8-bit | 40K | 134 tok/s | Not rated |
| Qwen2.5 7B Instruct | 8-bit | 32K | 146 tok/s | Not rated |
| Qwen2.5 VL 7B Instruct | 8-bit | 125K | 135 tok/s | Text and images |
| Qwen3 VL 4B Instruct | 8-bit | 176K | 207 tok/s | Text and images |
| Qwen3 4B | 8-bit | 40K | 223 tok/s | Not rated |
| Qwen3 32B | 6-bit | 16K | 44 tok/s | Not rated |
| Qwen3 4B Instruct 2507 | 8-bit | 176K | 223 tok/s | Not rated |
| Qwen3.6 27B | 8-bit | 24K | 41 tok/s | Text and images |
| Qwen2.5 Coder 14B Instruct | 8-bit | 32K | 75 tok/s | Coding |
| Qwen3 Coder 30B A3B Instruct | 6-bit | 64K | 207 tok/s | Coding |
| Qwen3.5 27B | 8-bit | 24K | 41 tok/s | Text and images |
| Qwen3.5 35B A3B | 6-bit | 96K | 227 tok/s | Text and images |
| NVIDIA Nemotron 3 Nano 30B A3B BF16 | 6-bit | 256K | 195 tok/s | Not rated |
| Qwen2.5 Coder 7B Instruct | 8-bit | 32K | 146 tok/s | Coding |
| Mistral 7B Instruct V0.2 | 8-bit | 32K | 143 tok/s | Not rated |
| Qwen3 30B A3B | 6-bit | 40K | 207 tok/s | Not rated |
| Qwen3 30B A3B Instruct 2507 | 6-bit | 64K | 207 tok/s | Not rated |
| Laguna XS.2 | 6-bit | 104K | 229 tok/s | Not rated |
| OTel 2.0 LLM 31B IT | 6-bit | 112K | 45 tok/s | Not rated |
| Qwen2.5 32B Instruct | 6-bit | 16K | 44 tok/s | Not rated |
| MiniCPM V 4.7 35B A3B | 6-bit | 120K | 227 tok/s | Text and images |
| Qwen2.5 14B Instruct | 8-bit | 32K | 75 tok/s | Not rated |
| Laguna XS 2.1 | 6-bit | 104K | 229 tok/s | Not rated |
| Diffusiongemma 26B A4B IT | 8-bit | 184K | 193 tok/s | Text and images |
| Qwen2.5 VL 32B Instruct | 5-bit | 28K | 49 tok/s | Text and images |
| Qwen2.5 Coder 32B Instruct | 6-bit | 16K | 44 tok/s | Coding |
| Qwen2 VL 7B Instruct | 8-bit | 32K | 135 tok/s | Text and images |
| Ornith 1.5 9B | 8-bit | 256K | 120 tok/s | Not rated |
| Qwen3 VL 30B A3B Instruct | 6-bit | 60K | 207 tok/s | Text and images |
| Mistral 7B V0.1 | 8-bit | 32K | 143 tok/s | Not rated |
| Meta Llama 3 8B Instruct Abliterated V3 | 8-bit | 8K | 137 tok/s | Not rated |
| EXAONE 3.5 7.8B Instruct | 8-bit | 32K | 138 tok/s | Not rated |
| Saiga Llama3 8B | 8-bit | 8K | 137 tok/s | Not rated |
| DeepSeek R1 0528 Qwen3 8B | 8-bit | 128K | 134 tok/s | Not rated |
| Gemma 3 12B IT | 8-bit | 128K | 92 tok/s | Text and images |
| Phi 3.5 Mini Instruct | 8-bit | 64K | 243 tok/s | Not rated |
| DeepSeek R1 Distill Qwen 32B | 6-bit | 16K | 44 tok/s | Not rated |
| Qwen3 4B Thinking 2507 | 8-bit | 176K | 223 tok/s | Not rated |
| Gemma 3 27B IT | 8-bit | 20K | 42 tok/s | Text and images |
| Phi 4 Mini Instruct | 8-bit | 128K | 238 tok/s | Not rated |
Fits, context and speed on the RTX 5090. A slower card with the same memory holds the same models at lower speed. Ranked by our use-case scores, then by Hugging Face downloads. The top 60 of 233 are shown.
Image and video models
| Model | Best fit | Per output | Strongest at |
|---|---|---|---|
| MiniMax H3 | 8-bit | 4.9 min | Video |
| LTX-2.5 | 8-bit | 47 s | Video |
| FLUX.1 [dev] | 16-bit | 9.5 s | Images |
| Qwen-Image | 8-bit | 2.7 min | Images |
| FLUX.2 [dev] | 6-bit | 85 s | Images |
| Qwen-Image 2.1 | 16-bit | 72 s | Images |
| Wan2.2 I2V A14B | 8-bit | 24 min | Video |
| FLUX.2 [klein] 9B | 16-bit | 3.0 s | Images |
| Wan2.1 I2V 14B | 8-bit | 24 min | Video |
| Wan 2.1 T2V-14B | 8-bit | 31 min | Video |
| Krea 2 Turbo | 16-bit | 9.4 s | Images |
| Stable Diffusion 3.5 Large | 16-bit | 38 s | Images |
| Qwen Image 2512 | 8-bit | 2.7 min | Images |
| Wan 2.2 A14B | 8-bit | 24 min | Video |
| HunyuanVideo 1.5 | 16-bit | 30 min | Video |
Machines with 32 GB
Questions
What is the best LLM for 32 GB of VRAM?
Of the open models we track, the strongest that run well on 32 GB are Qwen3.8 27B, Muse Glimmer 30B, Qwen3.6 35B A3B. 233 text models fit in all.
Calculated from published specs, not benchmarked. Every machine and tier