ji-farthing / MiMo-V2.5-Pro-DFlash-draft-GGUF-review-artifact

MiMo V2.5 Pro DFlash draft GGUF review artifact

This repository is an unofficial review artifact for an ik_llama follow-up PR. It contains a locally converted GGUF draft from XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash .

This is not a supported model release. I have not runtime-tested this GGUF with a MiMo V2.5 Pro target model. The intended use is reviewer/runtime-test convenience: an ik_llama maintainer or reviewer with suitable hardware can inspect or test the converted draft without repeating the conversion step.

Artifact status: the converted GGUF was uploaded on 2026-06-27.

File

• mimo-v25-pro-dflash-draft-bf16.gguf
• Size: 5,542,740,224 bytes
• SHA256: 756f04d4345c3396910a99e5de7849913c79e69f2186e59d136a03057dc43d3d

Provenance

• Source model repository: XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash
• Source subdirectory converted: dflash/
• Target metadata/tokenizer directory used for conversion: source repository root
• Output type: bf16
• Converter branch: filament/mimo-dflash-draft-convert
• Converter parent/base hash: f96eaddba8bed6a9a5e628bbf6a566775c70b49c
• Converter branch head hash: 910f53a5ebe42bfe0c707f7caec4924f37abb550
• Conversion date: 2026-06-27

Conversion shape observed locally:

• n_tensors = 63
• total_size = 5.5G
• sliding_window = 1024
• target_layer_ids = [0, 15, 31, 47, 69]
• five blk.N.attn_sinks.weight tensors

Local checks

• python3 -m py_compile convert_hf_to_gguf.py gguf-py/gguf/constants.py gguf-py/gguf/tensor_mapping.py
• cmake --build build --config Release -j$(nproc) --target llama-cli
• cmake --build build-cuda --config Release -j$(nproc) --target llama-server

The checks above cover converter output and build coverage for the optional attention-sink load/graph path. They do not prove MiMo V2.5 Pro runtime behavior.