Yes — DeepSeek-V4 Flash runs on Mac Studio 128GB
DeepSeek-V4 Flash (MoE 284B · 13B active) needs 76GB VRAM at IQ2 vs your 128GB unified. Estimated ~101+ tok/s (blazing).
284B MoE with 13B active and 1M context. Even aggressive UD-IQ2 needs ~75 GB combined — 128 GB unified or multi-GPU + fat RAM.
| quant | weights | + KV 8k | fits 128GB? |
|---|---|---|---|
| IQ2 | 76GB | ~76.3GB | ✓ yes |
| Q3 | 100GB | ~100.3GB | ✓ yes |
| Q4 | 160GB | ~160.3GB | × tight/no |
| Q8 | 290GB | ~290.3GB | × tight/no |
File size ≠ runtime. KV grows with context — halve to 4k/2k if long chats OOM. Leave 1–2GB free.
Can DeepSeek-V4 Flash run on Mac Studio 128GB?
DeepSeek-V4 Flash (MoE 284B · 13B active) needs 76GB VRAM at IQ2 vs your 128GB unified. Estimated ~101+ tok/s (blazing). Model context 1024k. Compare with all models or re-check with your exact specs on the bench.
Unload with ollama stop, cap layers --num-gpu 28, cut context to 2k. See the OOM fixer on the homepage.
Start IQ2. Bigger model at Q4 beats smaller at Q8. Only go Q8 if 2GB+ headroom remains.
About ~101+ tok/s (blazing) when 100% on GPU. Any CPU offload drops speed 5–20× — check ollama ps.
Yes — MIT allows commercial use.