DeepSeek-V4.1-Flash-1017, lossless MLX pack (for 512 GB Macs)
A bit-exact repack of deepseek-ai/DeepSeek-V4.1-Flash-1017
(model_type: deepseek_v41) in MLX safetensors layout: nothing is
re-quantized. The routed experts keep the release's FP4 (MXFP4), the trunk
keeps its FP8, and the MTP head and the vision tower ride along. Weight names
follow mlx-lm / oMLX, so the pack loads in both mlx-serve and oMLX.
This is the reference pack: it is what the quantized packs are scored against. It is for 512 GB Macs; on a 256 GB Mac use the quantized pack instead. We have not been able to load it on a 512 GB machine ourselves.
| size | |
|---|---|
| Resident (everything except Engram) | about 288 GB |
Engram tables (engram/, read from the SSD per token, not resident) |
about 189 GB |
| Download | about 477 GB |
Keep the model on the Mac's internal SSD: Engram on an external drive cost about 27% prefill in our measurements.
Run
mlx-serve --model ddalcu/DeepSeek-V4.1-Flash-1017-MLX-Serve-Lossless --serve
MLXServe.com for the app. For oMLX, point it at the
downloaded folder like any other DeepSeek-V4.1 pack. head and wo_a are
stored dense (bf16) because oMLX reads them that way.
Smaller packs
| pack | resident | use it if |
|---|---|---|
| this pack (lossless) | about 288 GB | 512 GB Mac, or as a reference |
| iQ-MLX-3.0bpw | 209 GB | 256 GB Mac |
License
MIT, as the base model.