What can the M3 Ultra 96 GB run?
| Text models: 112 | Visual models: 14 | Strongest first |
- Memory
- 96 GB unified
- Bandwidth
- 819 GB/s
- Maker
- Apple
Open the calculator on this machineApple M3 Ultra in our inventory
Text models that run on it
| Model | Best fit | Context | Speed | Strongest at |
|---|---|---|---|---|
| Qwen3.8 Flash-Next | 2-bit | 192K | 90 tok/s | Chat, Coding, Agents |
| Qwen3.8 27B | 8-bit | 256K | 17 tok/s | Chat, Coding, Agents, Docs |
| Muse Glimmer 30B | 8-bit | 128K | 16 tok/s | Docs |
| Mistral Medium 3.5 128B | 3-bit | 24K | 8.1 tok/s | Chat, Coding, Agents, Docs |
| Qwen3.5 122B A10B | 3-bit | 256K | 62 tok/s | Chat, Coding, Agents, Docs |
| Nemotron 3 Super 120B A12B | 3-bit | 256K | 42 tok/s | Chat, Coding, Agents, Docs |
| Nemotron 3.5 Bolt 74B A9B | 6-bit | 256K | 38 tok/s | Chat, Coding, Agents, Docs |
| Qwen3.6 35B A3B | 8-bit | 256K | 85 tok/s | Chat, Coding, Agents, Docs |
| Gemma 4 31B | 8-bit | 256K | 15 tok/s | Chat, Coding, Agents, Docs |
| Gemma 4 26B A4B | 8-bit | 256K | 85 tok/s | Chat, Coding, Agents, Docs |
| GLM-4.7 Flash | 8-bit | 198K | 80 tok/s | Chat, Coding, Agents |
| Mistral Small 4 119B | 4-bit | 16K | 80 tok/s | Chat, Docs |
| gpt-oss 120B | MXFP4 | 128K | 84 tok/s | Chat, Coding |
| Qwen3 Coder Next 80B | 6-bit | 256K | 87 tok/s | Coding, Agents |
| Nemotron 3.5 Lightning 30B A3B | 8-bit | 256K | 75 tok/s | Agents, Docs |
| Devstral Small 2 24B | FP8 | 256K | 20 tok/s | Coding |
| LFM2.5 24B A2B | 8-bit | 128K | 104 tok/s | Agents |
| Mistral Small 3.2 24B | 8-bit | 128K | 20 tok/s | Chat, Coding, Agents |
| Ministral 3 14B | 8-bit | 256K | 34 tok/s | Coding, Agents, Docs |
| Llama 3.3 70B | 6-bit | 44K | 8.8 tok/s | Chat, Coding, Agents, Docs |
| Olmo 3.1 32B | 8-bit | 64K | 15 tok/s | Chat, Coding, Agents, Docs |
| Qwen3 14B | 8-bit | 40K | 32 tok/s | Not rated |
| Qwen3 32B | 8-bit | 40K | 15 tok/s | Not rated |
| Qwen3.6 27B | 8-bit | 256K | 17 tok/s | Text and images |
| Qwen2.5 Coder 14B Instruct | 8-bit | 32K | 31 tok/s | Coding |
| Qwen3 Coder 30B A3B Instruct | 8-bit | 256K | 79 tok/s | Coding |
| Qwen3.5 27B | 8-bit | 256K | 17 tok/s | Text and images |
| Qwen3.5 35B A3B | 8-bit | 256K | 85 tok/s | Text and images |
| NVIDIA Nemotron 3 Nano 30B A3B BF16 | 8-bit | 256K | 75 tok/s | Not rated |
| GLM 4.5 Air | 4-bit | 28K | 43 tok/s | Not rated |
| Qwen3 30B A3B | 8-bit | 40K | 79 tok/s | Not rated |
| Qwen3 30B A3B Instruct 2507 | 8-bit | 256K | 79 tok/s | Not rated |
| Laguna XS.2 | 8-bit | 256K | 86 tok/s | Not rated |
| OTel 2.0 LLM 31B IT | 8-bit | 256K | 15 tok/s | Not rated |
| Qwen2.5 32B Instruct | 8-bit | 32K | 15 tok/s | Not rated |
| MiniCPM V 4.7 35B A3B | 8-bit | 256K | 85 tok/s | Text and images |
| Qwen2.5 14B Instruct | 8-bit | 32K | 31 tok/s | Not rated |
| Laguna XS 2.1 | 8-bit | 256K | 86 tok/s | Not rated |
| Diffusiongemma 26B A4B IT | 8-bit | 256K | 82 tok/s | Text and images |
| Qwen2.5 VL 32B Instruct | 8-bit | 125K | 14 tok/s | Text and images |
| Qwen2.5 Coder 32B Instruct | 8-bit | 32K | 15 tok/s | Coding |
| Laguna S 2.1 | 4-bit | 28K | 61 tok/s | Not rated |
| Qwen3 VL 30B A3B Instruct | 8-bit | 256K | 79 tok/s | Text and images |
| DeepSeek R1 Distill Qwen 32B | 8-bit | 128K | 15 tok/s | Not rated |
| Gemma 3 27B IT | 8-bit | 128K | 18 tok/s | Text and images |
| Ornith 1.5 35B A3B | 8-bit | 256K | 85 tok/s | Not rated |
| Ornith 1.5 122B | 3-bit | 256K | 53 tok/s | Not rated |
| GPT Neox 20B | 8-bit | 2K | 23 tok/s | Not rated |
| Phi 4 | 8-bit | 16K | 32 tok/s | Not rated |
| Qwen3 VL 32B Instruct | 8-bit | 136K | 14 tok/s | Text and images |
| Pixtral 12B | 8-bit | 352K | 37 tok/s | Text and images |
| DeepSeek R1 Distill Qwen 14B | 8-bit | 128K | 31 tok/s | Not rated |
| Qwen2.5 14B Instruct 1M | 8-bit | 256K | 31 tok/s | Not rated |
| Qwen2.5 72B Instruct | 6-bit | 32K | 8.6 tok/s | Not rated |
| Qwen3 Next 80B A3B Instruct | 6-bit | 232K | 87 tok/s | Not rated |
| Llama 4 Scout 17B 16E Instruct | 4-bit | 32K | 37 tok/s | Text and images |
| Llama Guard 4 12B | 8-bit | 288K | 39 tok/s | Text and images |
| Qwen2.5 VL 72B Instruct | 6-bit | 36K | 8.5 tok/s | Text and images |
| Qwen1.5 7B | 8-bit | 32K | 60 tok/s | Not rated |
| Solar Open 100B | 4-bit | 48K | 49 tok/s | Not rated |
Best fit is the highest quality, up to 8-bit, that leaves the use case’s context. Ranked by our use-case scores, then by Hugging Face downloads. Speeds are rough, plus or minus 25%. The top 60 of 112 are shown.
Image and video models
| Model | Best fit | Per output | Strongest at |
|---|---|---|---|
| MiniMax H3 | 8-bit | 35 min | Video |
| LTX-2.5 | 16-bit | 3.8 min | Video |
| FLUX.1 [dev] | 16-bit | 66 s | Images |
| Qwen-Image | 16-bit | 21 min | Images |
| FLUX.2 [dev] | 8-bit | 9.3 min | Images |
| Qwen-Image 2.1 | 16-bit | 9.1 min | Images |
| Wan2.2 I2V A14B | 16-bit | 3.1 h | Video |
| FLUX.2 [klein] 9B | 16-bit | 16 s | Images |
| Wan2.1 I2V 14B | 16-bit | 3.1 h | Video |
| Wan 2.1 T2V-14B | 16-bit | 3.9 h | Video |
| Krea 2 Turbo | 16-bit | 66 s | Images |
| Qwen Image 2512 | 16-bit | 21 min | Images |
| Wan 2.2 A14B | 16-bit | 3.1 h | Video |
| HunyuanVideo 1.5 | 16-bit | 3.8 h | Video |
Only just too big
- MiniMax M3
- Qwen3.5 397B A17B
- MiniMax M2.7
- MiniMax M2.5
- Laguna M.1
- Qwen3 VL 235B A22B Instruct
- Qwen3 235B A22B
- Ornith 1.5 397B
- Step 3.5 Flash
- MiniMax M2
- Qwen3 VL 235B A22B Thinking
- Qwen3 235B A22B Instruct 2507
- DeepSeek V2 Chat
- Le Triomphant ECE TW3
- TW3 JRGL V2
- Command A Plus 05 2026 Bf16
- MiniMax M2.1
- Qwen1.5 72B
- Intern S1
- Step 3.7 Flash
These need up to a quarter more memory than the M3 Ultra 96 GB has. Spilling into system RAM or a lower precision can get them running, slowly.
Questions
What is the best LLM for the M3 Ultra 96 GB?
Of the open models we track, the strongest that run well on the M3 Ultra 96 GB are Qwen3.8 Flash-Next, Qwen3.8 27B, Muse Glimmer 30B. 112 text models fit it in all.
Calculated from published specs, not benchmarked. All machines