422 Downloads Updated 2 days ago
ollama run maternion/mimo-v2.6-heretic:9b-instruct-q6_K
An abliterated variant of MiMo-V2.6-Distill-Qwen-9B. Refusal directions were identified and surgically removed from the model weights using heretic’s direction-ablation pipeline, reducing refusal behavior while preserving the base model’s reasoning, coding, and agentic capabilities.
| Property | Value |
|---|---|
| Base model | XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B |
| Method | Direction ablation (heretic) |
| Refusal prompts | LLM-LAT/harmful-dataset |
| Good prompts | mlabonne/harmless_alpaca |
| Trials | 200 (TPE, 60 random startup) |
| Best trial | 174 |
| Refusals | 20⁄100 |
| KL divergence | 0.0171 |
| Quantization | bnb_4bit |
Low KL divergence (0.0171) indicates the abliterated model remains numerically close to the base — reasoning and coding performance are effectively unchanged. The 20⁄100 refusal rate represents a significant reduction from the base model while avoiding the capability degradation seen at higher KL values.
Note: No model is fully “uncensored.” A small residual refusal rate remains on the most deeply-trained refusal categories. This is expected and preferable to the capability loss incurred by more aggressive ablation.
Run with Ollama:
#Thinking
ollama run maternion/mimo-v2.6-heretic:9b
#Instruct
ollama run maternion/mimo-v2.6-heretic:9b-instruct --think=false
| Tag | Precision | Params | Size |
|---|---|---|---|
9b |
4-bit (Q4_K_M alias) | thinking defaults | 5.6 GB |
9b-thinking |
4-bit (Q4_K_M) | temp 1.0, top_k 0 | 5.6 GB |
9b-instruct |
4-bit (Q4_K_M) | temp 0.6, top_k 20, presence 1.5 | 5.6 GB |
9b-thinking-q6_K |
6-bit (Q6_K) | temp 1.0, top_k 0 | 7.4 GB |
9b-instruct-q6_K |
6-bit (Q6_K) | temp 0.6, top_k 20, presence 1.5 | 7.4 GB |
9b-thinking-q8_0 |
8-bit (Q8_0) | temp 1.0, top_k 0 | 9.5 GB |
9b-instruct-q8_0 |
8-bit (Q8_0) | temp 0.6, top_k 20, presence 1.5 | 9.5 GB |
Thinking mode is the default (:9b / :9b-thinking); pass -instruct tags for non-thinking responses. The GGUFs carry a community-fixed chat template (JSON tool-call arguments + prefilled reasoning boundary) so tool calls and reasoning parsing work correctly in llama.cpp-based runtimes. No MTP/NextN tensors — converted with --no-mtp to avoid the upstream config mismatch.
MiMo-V2.6-Distill-Qwen-9B is a 9B agentic model developed by Xiaomi MiMo through supervised fine-tuning of Qwen3.5-9B on MiMo-generated data. It covers coding, general-purpose agent tasks, visual coding, and cybersecurity. We release this SFT checkpoint as a starting point for open research in agentic reinforcement learning.
Results for the released SFT checkpoint, as reported in the MiMo-V2.6 technical report.
| Domain | Benchmark | Metric | Qwen3.5-9B | MiMo-V2.6-Distill-Qwen-9B (SFT) |
|---|---|---|---|---|
| Code | SWE Verified | avg@3 | 60.0 | 61.1 |
| Code | SWE Pro | avg@3 | 32.0 | 44.6 |
| Code | MiMo Code (mini)† | avg@3 | 19.5 | 51.6 |
| Cyber | MiMo Cyber (mini)† | avg@3 | 5.7 | 31.3 |
| General | AutomationBench v1.0.6 | avg@1 | 5.0 | 30.3 |
| General | Terminal Bench 2.1 | avg@1 | 27.0 | 37.1 |
| General | Toolathlon-Verified | avg@1 | 25.9 | 35.2 |
| General | OfficeQA | avg@1 | 9.0 | 19.5 |
| General | JobBench | avg@1 | 2.6 | 18.3 |
| General | MiMo General (mini)† | avg@1 | 28.5 | 62.2 |
| Visual | MiMo Visual Coding (mini)† | avg@1 | 61.7 | 64.0 |
† Internal evaluation sets.
The weighted SFT data mixture contains 77.4B total tokens, including 27.2B loss-bearing tokens.
| Domain | Total tokens (B) | Token share (%) | Loss-bearing tokens (B) |
|---|---|---|---|
| Code | 23.2 | 29.9 | 7.3 |
| Cyber | 11.0 | 14.2 | 4.8 |
| General | 22.0 | 28.5 | 5.7 |
| Visual | 21.2 | 27.4 | 9.4 |
| Total | 77.4 | 100.0 | 27.2 |
MIT license. See the upstream repository for details.
@misc{mimo2026v26,
title={MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement},
author={{Xiaomi MiMo Team}},
year={2026},
howpublished={\url{https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL}},
}