Entrit/DeepSeek-V4-Flash-d2swap-mp2

DeepSeek-V4-Flash d2 experts

DeepSeek-V4-Flash with its routed experts quantized by the Entrit d2 codec.

What changed from the base model

All 256 routed experts are replaced by d2 values. Each weight is one of 9 levels, {0, ±1, ±2.83, ±5.20, ±8}, times one scale per group of 64 weights: about 3.4 bits of information per weight. Everything else (attention, shared expert, router, embedding, head) is the base model's FP4 release, unchanged.

Use

The files are two tensor-parallel shards (model0-mp2, model1-mp2) in the layout of DeepSeek's reference V4 inference code. Run them with that code at tensor-parallel 2, passing the config.json in this repo. They do not load in transformers or vLLM. The checkpoint is 163 GB.

Results

GSM8K 0.97, MMLU 0.95, MATH-500 0.95: about 99% of the FP4 base model on the same harness.

License and notice

MIT License, inherited from deepseek-ai/DeepSeek-V4-Flash; see LICENSE. This is a modified version of that model: the routed experts were quantized as described above.

The d2 codec is the subject of pending patent applications by Entrit Systems. Contact: eric@entrit.io