256 GB unified memory and 819 GB/s of memory bandwidth. The strongest single-device fits at 4-bit are Inkling Small (~910 tok/s), MiniMax M2.5 (~149 tok/s), Gemma 4 31B (~52 tok/s), Qwen3 235B (Thinking) (~74 tok/s) and Qwen3 32B (~50 tok/s).
72 open-weight models fit in Mac Studio M3 Ultra 256 GB at 4-bit, ranked by Analyzer Index.
| Model | Index | Memory | Est. tok/s | API $/month | Verdict |
|---|---|---|---|---|---|
| Inkling Small Thinking Machines266B · 1.8B active | 62.2 | 160 GB | ~910 | $75.00 | pays back in 75 mo |
| MiniMax M2.5 MiniMax228.7B · 11B active | 57.8 | 137 GB | ~149 | $50.75 | pays back in 110 mo |
| Gemma 4 31B Google31.3B | 57.4 | 19 GB | ~52.3 | $18.50 | pays back in 303 mo |
| Qwen3 235B (Thinking) Alibaba (Qwen)235.1B · 22B active | 51.6 | 141 GB | ~74.5 | $133 | pays back in 42 mo |
| Qwen3 32B Alibaba (Qwen)32.8B | 50.0 | 20 GB | ~49.9 | $15.00 | pays back in 373 mo |
| gpt-oss-120b OpenAI116.8B · 5.7B active | 45.8 | 70 GB | ~287 | $6.75 | pays back in 830 mo |
| Qwen3 235B Alibaba (Qwen)235.1B · 22B active | 44.1 | 141 GB | ~74.5 | $17.50 | pays back in 320 mo |
| QwQ 32B Alibaba (Qwen)32.8B | 37.2 | 20 GB | ~49.9 | — | no API price |
| gpt-oss-20b OpenAI20.9B · 4.2B active | 35.2 | 13 GB | ~390 | $4.05 | pays back in 1383 mo |
| Gemma 3 27B Google27.4B | 28.5 | 16 GB | ~59.8 | $12.00 | pays back in 467 mo |
| Mistral Large 2 Mistral122.6B | 26.8 | 74 GB | ~13.4 | $350 | pays back in 16 mo |
| Llama 3.1 70B Instruct Meta70.6B | 24.3 | 42 GB | ~23.2 | $50.00 | pays back in 112 mo |
| Llama 3.3 70B Instruct Meta70.6B | 20.9 | 42 GB | ~23.2 | $18.00 | pays back in 311 mo |
| Llama 4 Scout Meta108.6B · 17B active | 12.3 | 65 GB | ~96.4 | $14.00 | pays back in 400 mo |
| Nemotron 3.5 Lightning NVIDIA31.6B · 3B active | — | 19 GB | ~546 | $5.50 | pays back in 1018 mo |
| Qwen3.6 35B A3B Alibaba (Qwen)36B · 3B active | — | 22 GB | ~546 | $22.50 | pays back in 249 mo |
| Qwen3.5-35B-A3B Alibaba (Qwen)36B · 3B active | — | 22 GB | ~546 | $25.25 | pays back in 222 mo |
| Nemotron 3 Nano 30B A3B NVIDIA31.6B · 3B active | — | 19 GB | ~546 | $10.00 | pays back in 560 mo |
| Qwen3 VL 30B A3B Thinking Alibaba (Qwen)31.1B · 3B active | — | 19 GB | ~546 | $54.00 | pays back in 104 mo |
| Qwen3 VL 30B A3B Instruct Alibaba (Qwen)31.1B · 3B active | — | 19 GB | ~546 | $30.00 | pays back in 187 mo |
| Qwen3 Next 80B A3B Thinking Alibaba (Qwen)81.3B · 3B active | — | 49 GB | ~546 | $45.00 | pays back in 124 mo |
| Qwen3 Next 80B A3B Instruct Alibaba (Qwen)81.3B · 3B active | — | 49 GB | ~546 | $29.20 | pays back in 192 mo |
| Qwen3 Coder 30B A3B Instruct Alibaba (Qwen)30.5B · 3B active | — | 18 GB | ~546 | $14.00 | pays back in 400 mo |
| Qwen3 30B A3B Instruct 2507 Alibaba (Qwen)30.5B · 3B active | — | 18 GB | ~546 | $16.50 | pays back in 339 mo |
| Hunyuan A13B Instruct Tencent (Hunyuan)80.4B · 3B active | — | 48 GB | ~546 | $28.25 | pays back in 198 mo |
| Qwen3 30B A3B Alibaba (Qwen)30.5B · 3B active | — | 18 GB | ~546 | $24.50 | pays back in 229 mo |
| Ministral 3 3B 2512 Mistral3.8B | — | 2 GB | ~431 | $12.50 | pays back in 448 mo |
| Qwen3 Coder Next Alibaba (Qwen)79.7B · 3.9B active | — | 48 GB | ~420 | $32.00 | pays back in 175 mo |
| Gemma 4 26B A4B Google25.8B · 4B active | — | 16 GB | ~410 | $9.70 | pays back in 577 mo |
| Gemma 3 4B Google4.3B | — | 3 GB | ~381 | $5.25 | pays back in 1067 mo |
| GLM 4.7 Flash Zhipu AI (GLM)31.2B · 4.6B active | — | 19 GB | ~356 | $16.05 | pays back in 349 mo |
| Mistral Small 4 Mistral119.4B · 7.1B active | — | 72 GB | ~231 | $30.00 | pays back in 187 mo |
| Stable Diffusion 3.5 Large Stability AI8.1B | — | 5 GB | ~202 | — | no API price |
| Qwen3 VL 8B Instruct Alibaba (Qwen)8.8B | — | 5 GB | ~186 | $43.75 | pays back in 128 mo |
| Ministral 3 8B 2512 Mistral8.9B | — | 5 GB | ~184 | $18.75 | pays back in 299 mo |
| Qwen3.5-9B Alibaba (Qwen)9.7B | — | 6 GB | ~169 | $11.25 | pays back in 498 mo |
| Qwen3.5-122B-A10B Alibaba (Qwen)125.1B · 10B active | — | 75 GB | ~164 | $78.00 | pays back in 72 mo |
| MiniMax M2.7 MiniMax228.7B · 11B active | — | 137 GB | ~149 | $42.00 | pays back in 133 mo |
| MiniMax M2.1 MiniMax228.7B · 11B active | — | 137 GB | ~149 | $60.00 | pays back in 93 mo |
| MiniMax M2 MiniMax228.7B · 11B active | — | 137 GB | ~149 | $51.00 | pays back in 110 mo |
| MiMo-V2.5 Xiaomi (MiMo)310.8B · 11.2B active | — | 187 GB | ~146 | $17.85 | pays back in 314 mo |
| FLUX.1 [dev] Black Forest Labs11.9B | — | 7 GB | ~138 | — | no API price |
| Nemotron 3 Super NVIDIA123.6B · 12B active | — | 74 GB | ~137 | $18.50 | pays back in 303 mo |
| Gemma 3 12B Google12.2B | — | 7 GB | ~134 | $7.00 | pays back in 800 mo |
| Ministral 3 14B 2512 Mistral13.9B | — | 8 GB | ~118 | $25.00 | pays back in 224 mo |
| DeepSeek V4 Flash 0423 DeepSeek290.9B · 14.1B active | — | 175 GB | ~116 | $3.34 | pays back in 1675 mo |
| GLM 4.6V Zhipu AI (GLM)107.7B · 14.3B active | — | 65 GB | ~115 | $52.50 | pays back in 107 mo |
| GLM 4.5V Zhipu AI (GLM)107.7B · 14.3B active | — | 65 GB | ~115 | $105 | pays back in 53 mo |
| Qwen3 14B Alibaba (Qwen)14.8B | — | 9 GB | ~111 | $15.50 | pays back in 361 mo |
| GLM 4.5 Air Zhipu AI (GLM)110.5B · 15B active | — | 66 GB | ~109 | $34.25 | pays back in 164 mo |
| DeepSeek V4 Flash Vision Exp DeepSeek304.6B · 15.2B active | — | 183 GB | ~108 | $37.73 | pays back in 148 mo |
| Hy3 Tencent (Hunyuan)298.8B · 21B active | — | 179 GB | ~78.0 | $26.40 | pays back in 212 mo |
| Qwen3 VL 235B A22B Instruct Alibaba (Qwen)235.7B · 22B active | — | 141 GB | ~74.5 | $42.00 | pays back in 133 mo |
| Mistral Small 3 Mistral23.6B | — | 14 GB | ~69.4 | $7.00 | pays back in 800 mo |
| Mistral Small 3.2 24B Mistral24B | — | 14 GB | ~68.3 | $12.50 | pays back in 448 mo |
| Mistral Small 3.1 24B Mistral24B | — | 14 GB | ~68.3 | $48.97 | pays back in 114 mo |
| Voxtral Small 24B 2507 Mistral24.3B | — | 15 GB | ~67.4 | $17.50 | pays back in 320 mo |
| DeepSeek V4 Flash 0731 DeepSeek304.2B · 27.4B active | — | 183 GB | ~59.8 | $7.77 | pays back in 720 mo |
| Qwen3.8 27B Alibaba (Qwen)27.8B | — | 17 GB | ~58.9 | $14.50 | pays back in 386 mo |
| Qwen3.6 27B Alibaba (Qwen)27.8B | — | 17 GB | ~58.9 | $70.00 | pays back in 80 mo |
| Qwen3.5-27B Alibaba (Qwen)27.8B | — | 17 GB | ~58.9 | $58.50 | pays back in 96 mo |
| Muse Glimmer 30B Meta29.8B | — | 18 GB | ~55.0 | $60.00 | pays back in 93 mo |
| Mixtral 8x22B Mistral140.6B · 44B active | — | 84 GB | ~37.2 | $350 | pays back in 16 mo |
| Qwen3.8 Flash Alibaba (Qwen)180B · 61.6B active | — | 108 GB | ~26.6 | $26.75 | pays back in 209 mo |
| Qwen2.5 72B Instruct Alibaba (Qwen)72.7B | — | 44 GB | ~22.5 | $46.00 | pays back in 122 mo |
| Qwen2.5 VL 72B Instruct Alibaba (Qwen)73.4B | — | 44 GB | ~22.3 | $105 | pays back in 53 mo |
| Command A Cohere111.1B | — | 67 GB | ~14.7 | $500 | pays back in 11 mo |
| Devstral 2 2512 Mistral125B | — | 75 GB | ~13.1 | $90.00 | pays back in 62 mo |
| Step 3.5 Flash StepFun199.4B | — | 120 GB | ~8.2 | $17.50 | pays back in 320 mo |
| Step 3.7 Flash StepFun201.4B | — | 121 GB | ~8.1 | $48.75 | pays back in 115 mo |
| MiniMax M3.1 Flash MiniMax220.5B | — | 132 GB | ~7.4 | $21.50 | pays back in 260 mo |
| MiMo-V2.6-Flash Xiaomi (MiMo)310.8B | — | 187 GB | ~5.3 | $13.65 | pays back in 410 mo |
Memory need is weights × 1.2 for runtime overhead. Speed is a theoretical single-stream ceiling (memory bandwidth ÷ active-weight bytes); real engines usually reach 50–80% of it, and batching many requests raises total throughput well beyond one stream. Rental and purchase prices are typical figures checked 2026-10-14; power and engineering time are not included.