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

A decensored variant of XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B, produced with Heretic v1.4.0 (directional ablation / "abliteration"). Xiaomi's MiMo-V2.6 distill is a 9B agentic model covering coding, visual coding, tool use, and security work; refusal behaviour is suppressed here via targeted weight edits to the attention output and MLP down-projections rather than fine-tuning, so the SFT'd agentic behaviour and the vision encoder are left largely intact.

Who this is for: developers who want an uncensored agentic coding model - this is the one in this collection tuned for SWE-bench-style work, terminal and tool use, and visual coding, and the refusals here dropped from 99/100 to 8/100, the lowest absolute refusal count of this batch. MIT licensed, multimodal, 262K context, and it runs locally via GGUF on a 16 GB GPU. Best for local coding agents, code review, and security research workflows.

Runs on your gaming PC

Full GGUF ladder included - pick the quant that fits your card:

Weights only, at this model's native 9B size; add ~1 GB per 32K of context. OOM? Drop one quant level. Headroom to spare? Go one up.

Abliteration parameters

Trial 109 of a 200-trial Heretic run (seed 1903262967).

Performance

KL divergence of 0.0477 combined with refusals at 8/100 makes this the best refusal-removal / fidelity trade-off in this batch: the harmful evaluation set drops from 99/100 to 8/100 while the SFT'd agentic behaviour stays close to the original. Note direction_index is a single index (23.77) here rather than per-layer, which is why the edit lands tighter than the per-layer runs.

Why abliteration instead of fine-tuning

Fine-tuning a "helpful" persona on top of RLHF'd refusals fights the base model's training and tends to degrade coherence. Abliteration instead finds and edits the specific weight directions responsible for refusal, leaving the rest of the network (and its capabilities) untouched. See the Heretic repo and the original abliteration writeup for the mechanism.

Made with ❤️ by RACER IS OP (saidutta69) — follow for more uncensored models

Files

Safetensors

BF16, ~9B. The reproduce/ directory carries the full Heretic recipe - config.toml, requirements.txt, the Optuna study journal, and SHA-256 sums - so this exact model can be regenerated bit-for-bit. Reproduce it with heretic --reproduce reproduce/reproduce.json.

GGUF quantizations

Full quantization set (F16 + Q4_K_M, Q5_K_M, Q6_K, Q8_0) produced with llama.cpp.

Qwen3.5 architecture (qwen35) - loads natively in llama.cpp / Ollama / LM Studio / Jan.

Run llama serve -hf saidutta69/MiMo-V2.6-Distill-Qwen-9B-heretic to pull the default quant.

Quickstart

# llama.cpp
llama serve -hf saidutta69/MiMo-V2.6-Distill-Qwen-9B-heretic

# transformers
from transformers import AutoProcessor, AutoModelForImageTextToText

model_name = "saidutta69/MiMo-V2.6-Distill-Qwen-9B-heretic"
model = AutoModelForImageTextToText.from_pretrained(model_name, dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_name)

messages = [{"role": "user", "content": "Find the bug in this function and rewrite it:\n\ndef avg(xs):\n    return sum(xs)/len(xs)+1"}]
inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True,
                                        return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=2048)
print(processor.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Also runnable via Ollama, LM Studio, Jan, vLLM, SGLang.

Visual coding

The model keeps its vision encoder, so you can hand it a screenshot and ask for the code that produced the UI:

messages = [{"role": "user", "content": [
    {"type": "image", "image": "https://example.com/screenshot.png"},
    {"type": "text", "text": "Write the React and Tailwind that reproduces this layout."},
]}]

Tool use

Pass tool definitions straight through apply_chat_template — MiMo-V2.6 was SFT'd for tool calling, so the format is already in-distribution and does not need adapting.

Model details

Base model evaluation

As reported in the MiMo-V2.6 technical report for the released SFT checkpoint:

FAQ

What is MiMo-V2.6-Distill-Qwen-9B-heretic?

MiMo-V2.6-Distill-Qwen-9B-heretic is a decensored build of XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B produced with Heretic 1.4.0 (directional ablation). It is the same weights with the refusal direction removed, not a new fine-tune, so its capabilities track the base model.

Is MiMo-V2.6-Distill-Qwen-9B-heretic actually uncensored?

Yes. On Heretic's 100-prompt refusal harness it drops from 99/100 to 8/100 at KL divergence 0.0477 from the base model. These are measured values, not a claim; reproduction data is in reproduce/.

How do I run MiMo-V2.6-Distill-Qwen-9B-heretic locally?

Text only: llama serve -hf saidutta69/MiMo-V2.6-Distill-Qwen-9B-heretic:Q4_K_M (one of 5 GGUF quants). The GGUFs drop the vision tower; use the BF16 safetensors plus AutoProcessor for image input.

Responsible use

Refusal suppression is deliberate and works as intended: this model will comply with requests the base model would refuse, including some it shouldn't. There is no safety filtering layered on top. This model is explicitly tuned for cybersecurity work, so take that seriously: you are responsible for how you deploy it and for what you point it at. It inherits MiMo-V2.6's factual limitations and biases; abliteration removes refusal directions, it doesn't add capability or judgment.

License

Inherits the mit license from the base model.

Related

• XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B — the base model
• Qwen/Qwen3.5-9B — the upstream backbone
• RACER IS OP — Heretic Models — full collection

Abliteration hardware

Recorded in reproduce/reproduce.json:

• GPU: NVIDIA GeForce RTX 5090 (31.36 GB VRAM)
• CPU: AMD Ryzen 7 9700X 8-Core Processor
• OS: Linux-6.8.0-142-generic-x86_64-with-glibc2.35
• CUDA / driver: CUDA 13.0, driver 595.84
• Software: PyTorch 2.14.1+cu130, Heretic 1.4.0

Heretic notes that KL-divergence and refusal numbers can vary slightly with hardware and library versions.