MiMo-V2.6 Distill Qwen 9B DWM a1p1 — Q3_K_M GGUF

This repository contains a Q3_K_M GGUF conversion of a DWM-modified MiMo-V2.6 Distill Qwen 9B checkpoint, along with the BF16 vision projector required for multimodal use.

Files

The intermediate BF16 main-model GGUF is intentionally not included.

Conversion and validation

• Quantization: Q3_K_M (4.12 effective bits per weight)
• Architecture metadata: qwen35
• Decoder blocks: 32
• Trained context length: 262,144 tokens
• Main-model tensors: 427
• Vision-projector tensors: 334
• llama.cpp commit: d1a92352cbd417fd840b4e765c0b82f5fe3d1d89

The source configuration advertised one MTP/NextN layer but contained no serialized mtp.* tensors. The corrected conversion therefore excluded the absent NextN layer. The main model and vision projector were load-tested together, and a short text-generation smoke test completed successfully.

Usage

Use a recent llama.cpp build with MiMo/Qwen3.5 multimodal support:

llama-server \
  -m MiMo-V2.6-Distill-Qwen-9B-DWM-a1p1-Q3_K_M.gguf \
  --mmproj mmproj-MiMo-V2.6-Distill-Qwen-9B-DWM-a1p1-BF16.gguf \
  --ctx-size 262144 \
  --jinja

Actual usable context depends on available memory. File hashes are recorded in SHA256SUMS.

Access

This is a public, manually gated research release. Submit an access request through the Hugging Face repository page.