What can the M5 Ultra 512 GB run?
| Text models: 69 | Visual models: 0 | Strongest first |
- Memory
- 512 GB unified
- Bandwidth
- 1,200 GB/s
- Maker
- Apple
Text models that run on it
| Model | Best fit | Context | Speed | Strongest at |
|---|---|---|---|---|
| MiMo V2.6 Pro | 2-bit | 128K | 33 tok/s | Chat, Coding, Agents, Docs |
| DeepSeek V4.1 Flash | 3-bit | 1M | 62 tok/s | Chat, Coding, Agents, Docs |
| GLM-5.3 | 3-bit | 216K | 28 tok/s | Chat, Coding, Agents, Docs |
| GLM-5.3 Flash | FP8 | 1M | 33 tok/s | Chat, Coding, Agents, Docs |
| Qwen3.8 Flash-Next | 8-bit | 256K | 66 tok/s | Chat, Coding, Agents |
| Muse Spark Open 816B A42B | 2-bit | 1M | 35 tok/s | Chat, Docs |
| MiMo V2.6 Flash | FP4 + FP8 | 1M | 49 tok/s | Coding, Agents |
| MiniMax M3 | 6-bit | 320K | 31 tok/s | Docs |
| DeepSeek V4 Flash | FP4 + FP8 | 1M | 33 tok/s | Docs |
| DeepSeek V4 Pro | 1.6-bit | 1M | 47 tok/s | Chat, Coding, Agents, Docs |
| Kimi K2.6 | 2-bit | 80K | 42 tok/s | Chat, Coding, Agents, Docs |
| Hy3 | 8-bit | 224K | 25 tok/s | Chat, Coding, Agents, Docs |
| Nemotron 3 Ultra 550B A55B | 4-bit | 256K | 17 tok/s | Chat, Coding, Docs |
| Qwen3.5 397B A17B | 6-bit | 256K | 38 tok/s | Agents, Docs |
| MiniMax M2.7 | FP8 | 200K | 46 tok/s | Coding, Agents |
| DeepSeek V3.2 | 3-bit | 160K | 32 tok/s | Chat, Coding, Agents, Docs |
| Mistral Medium 3.5 128B | FP8 | 256K | 5.7 tok/s | Chat, Coding, Agents, Docs |
| Qwen3.5 122B A10B | 8-bit | 256K | 49 tok/s | Chat, Coding, Agents, Docs |
| Nemotron 3 Super 120B A12B | 8-bit | 256K | 35 tok/s | Chat, Coding, Agents, Docs |
| Mistral Small 4 119B | FP8 | 256K | 74 tok/s | Chat, Docs |
| Inkling | 2-bit | 4K | 35 tok/s | Text and images |
| DeepSeek R1 | 3-bit | 160K | 30 tok/s | Not rated |
| DeepSeek V3 | 3-bit | 160K | 30 tok/s | Not rated |
| GLM 5 | 3-bit | 198K | 27 tok/s | Not rated |
| GLM 4.5 Air | 8-bit | 128K | 37 tok/s | Not rated |
| Inkling Small | 8-bit | 4K | 47 tok/s | Text and images |
| MiniMax M2.5 | FP8 | 192K | 42 tok/s | Not rated |
| GLM 5.1 | 3-bit | 198K | 27 tok/s | Not rated |
| Laguna M.1 | 8-bit | 256K | 24 tok/s | Not rated |
| Laguna S 2.1 | 8-bit | 1M | 55 tok/s | Not rated |
| Qwen3 VL 235B A22B Instruct | 8-bit | 256K | 23 tok/s | Text and images |
| GLM 5.2 | 3-bit | 216K | 27 tok/s | Not rated |
| Qwen3 235B A22B | 8-bit | 40K | 23 tok/s | Not rated |
| Ornith 1.5 122B | 8-bit | 256K | 41 tok/s | Not rated |
| Kimi K2 Instruct | 2-bit | 80K | 35 tok/s | Not rated |
| Ornith 1.5 397B | 6-bit | 256K | 29 tok/s | Not rated |
| Ornith 1.0 397B | 6-bit | 256K | 37 tok/s | Not rated |
| Step 3.5 Flash | 8-bit | 256K | 52 tok/s | Not rated |
| DeepSeek V3 0324 | 3-bit | 160K | 30 tok/s | Not rated |
| DeepSeek R1 0528 | 3-bit | 160K | 30 tok/s | Not rated |
| GLM 4.5 | 8-bit | 24K | 16 tok/s | Not rated |
| Llama 4 Scout 17B 16E Instruct | 8-bit | 1.3M | 32 tok/s | Text and images |
| MiniMax M2 | FP8 | 192K | 42 tok/s | Not rated |
| MiMo V2.5 | FP8 | 1M | 34 tok/s | Not rated |
| Solar Open 100B | 8-bit | 128K | 43 tok/s | Not rated |
| GLM 5.3 Flash BF16 | 8-bit | 1M | 32 tok/s | Text and images |
| Qwen3 VL 235B A22B Thinking | 8-bit | 256K | 23 tok/s | Text and images |
| Qwen3 235B A22B Instruct 2507 | 8-bit | 256K | 23 tok/s | Not rated |
| NVIDIA Nemotron 3.5 Super VL 120B A12B BF16 | 8-bit | 1M | 35 tok/s | Text and images |
| Hy3 Preview | 8-bit | 216K | 25 tok/s | Not rated |
| Qwen3 Coder 480B A35B Instruct | 5-bit | 184K | 25 tok/s | Coding |
| GLM 4.7 | 8-bit | 24K | 16 tok/s | Not rated |
| GLM 5.3 BF16 | 3-bit | 216K | 27 tok/s | Not rated |
| GLM 4.6 | 8-bit | 28K | 17 tok/s | Not rated |
| DeepSeek V2 Chat | 8-bit | 160K | 27 tok/s | Not rated |
| MiMo V2 Flash | FP8 | 256K | 36 tok/s | Not rated |
| Le Triomphant ECE TW3 | 8-bit | 32K | 9.7 tok/s | Not rated |
| TW3 JRGL V2 | 8-bit | 32K | 9.7 tok/s | Not rated |
| Command A Plus 05 2026 Bf16 | 8-bit | 195K | 26 tok/s | Text and images |
| GLM 4.5V | 8-bit | 64K | 39 tok/s | Text and images |
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 69 are shown.
Only just too big
These need up to a quarter more memory than the M5 Ultra 512 GB has. Spilling into system RAM or a lower precision can get them running, slowly.
Questions
What is the best LLM for the M5 Ultra 512 GB?
Of the open models we track, the strongest that run well on the M5 Ultra 512 GB are MiMo V2.6 Pro, DeepSeek V4.1 Flash, GLM-5.3. 69 text models fit it in all.
Calculated from published specs, not benchmarked. All machines