puwaer/DeepSeek-V4-Flash-0731-ream-150b

Fix inference/config.json: n_routed_experts and n_mtp_layers were copied verbatim from the base model's inference/config.json (256 experts, 3 MTP layers) instead of reflecting this checkpoint's actual compressed architecture (132 experts, 0 MTP layers -- the MTP modules were removed during compression, confirmed against model.safetensors.index.json). inference/model.py uses these fields to size the expert gate and MTP prediction heads, so the stale values would have caused a shape mismatch loading this checkpoint through the bundled reference inference/ scripts.

Upload 8 files

Upload folder using huggingface_hub

initial commit