MiMo-V2.6-Distill-Qwen-9B-Decision

A general decision adapter for XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B: supply a state (evidence), questions and allowed options; receive probabilities over the options in one forward pass, without generating text. Rank-16 LoRA adapter plus a 256-d pointer head, run with the Kev runtime (commit 0a6d3faf). Not a standalone model.

It is the starting point of airun-ai/GlanceRAG-MiMo-9B-Reranker and one of the three modes served from one backbone in the Glance paper (Adilet Metinov, Chingiz Arziev; Airun Intelligence Labs, 2026; arXiv:2610.40454).

Results

• JevBench public 231 (self-run, pinned 7625ecb1): 184/231 (easy 48/48, original 71/72, hard 65/111), bf16 on an RTX PRO 6000.
• Pilot test (520 questions): accuracy 83.85% vs 75.19% for the base model's restricted-option logits.
• As a zero-shot reranker (Glance one pass, BM25 top-20): SciFact nDCG@10 0.715, NFCorpus 0.341, vs bge-reranker-v2-m3 0.721 / 0.326.

Use

python -m pip install -r requirements.txt
python -m pip install --no-deps triton==3.8.0
python inference.py --request request.json

{"state":"The customer was charged twice for an invoice.","questions":{"route":{"type":"choice","instructions":"Choose the appropriate support team.","criteria":{"billing":"Payments and invoices","technical":"Software problems","sales":"New purchases"}}}}

Supports choice, noul (yes/no) and ordinal score questions. Default temperature 1.0.

Citation

@misc{metinov2026glance,
  title  = {Glance: The Generator Is Its Own Reranker},
  author = {Metinov, Adilet and Arziev, Chingiz},
  year          = {2026},
  eprint        = {2610.40454},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2610.40454}
}

License: Apache-2.0. See LICENSE and NOTICE.