base_model: XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B
library_name: mlc-llm
tags:

• mlc-llm
• web-llm
• webgpu
• q4f16_1
• mimo
• qwen3.5
• abliterated
• uncensored
• agent
• coding
  license: mit

MiMo-V2.6-Distill-Qwen-9B-abliterated-q4f16_1-MLC

MLC-compiled weights for XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B (Abliterated / Uncensored version), quantized in q4f16_1 for 100% in-browser WebLLM and WebGPU inference.

🌟 Overview & What Makes This Special

MiMo-V2.6-Distill-Qwen-9B is Xiaomi's reinforcement-learning distilled reasoning model based on Qwen3.5-9B, achieving top-tier open-source coding performance (SWE-bench Verified 61.1%, SWE-bench Pro 44.6%).

This repository hosts an Abliterated (Refusal-Free) conversion in MLC format, specifically designed for local coding agents (like Broub) and local creative applications where ethical-filter refusals ("I cannot assist with this request...") obstruct valid cybersecurity research, penetration testing, malware reverse engineering, and creative writing.

🔬 Creator Insights & Technical Findings ()

1. Zero-Compile WebGPU Wasm Reuse ()

• Architecture: Qwen3_5ForConditionalGeneration (model_type: qwen3_5).
• Layer count (32), hidden size (4096), intermediate size (12288), and vocabulary (248320) are 100% identical to official Qwen3.5-9B.
• Wasm: You can directly reuse the official prebuilt Qwen3.5-9B-q4f16_1_cs1k-webgpu.wasm from mlc-ai without compiling any C++ or WebAssembly yourself!

2. Surgical Abliteration on Hybrid Architecture ()

Qwen3.5 is not a standard Transformer; it features a unique hybrid attention design:

• Every 4th layer is a standard Full Attention (self_attn with o_proj).
• The other 3 layers are Linear Attention (Gated DeltaNet / Mamba with linear_attn.out_proj).
• All layers feature an MLP (mlp.down_proj).

To eliminate refusals without damaging coding or logical reasoning:

• Refusal direction vectors $\vec{r} \in \mathbb{R}^{4096}$ were extracted from the residual streams of contrastive prompt pairs.
• Instead of wiping all layers, we applied surgical orthogonal projection subtraction restricted strictly to intermediate layers (Layer 8 to 26):
  $$W_{\text{new}} \leftarrow W - \vec{r} \otimes (\vec{r}^T W)$$
  where $W$ is the output projection (o_proj, out_proj, or down_proj).
• This mathematical guarantee ensures $(W_{\text{new}} x) \cdot \vec{r} \approx 0$, removing refusal steering while preserving 100% of the token syntax and reasoning capabilities.

📊 Evaluation & Verification ()

Evaluation was performed directly prior to quantization (see ablation_benchmark_report.json in this repo):

🚀 Usage in Broub & WebLLM

Broub (Browser-Local Coding Agent)

Go to the Models page in Broub, select Custom Model, and enter:

{
  "model": "https://huggingface.co/aaaaaeeeee/MiMo-V2.6-Distill-Qwen-9B-abliterated-q4f16_1-MLC",
  "model_id": "MiMo-V2.6-Distill-Qwen-9B-abliterated-q4f16_1-MLC",
  "model_lib": "https://raw.githubusercontent.com/mlc-ai/binary-mlc-llm-libs/main/web-llm-models/v0_2_84/base/Qwen3.5-9B-q4f16_1_cs1k-webgpu.wasm",
  "vram_required_MB": 6433,
  "overrides": {
    "context_window_size": 4096,
    "max_history_size": 1
  }
}

WebLLM JavaScript API

import { CreateWebWorkerMLCEngine } from "@mlc-ai/web-llm";

const appConfig = {
  model_list: [
    {
      model: "https://huggingface.co/aaaaaeeeee/MiMo-V2.6-Distill-Qwen-9B-abliterated-q4f16_1-MLC",
      model_id: "MiMo-V2.6-Distill-Qwen-9B-abliterated-q4f16_1-MLC",
      model_lib: "https://raw.githubusercontent.com/mlc-ai/binary-mlc-llm-libs/main/web-llm-models/v0_2_84/base/Qwen3.5-9B-q4f16_1_cs1k-webgpu.wasm",
      vram_required_MB: 6433
    }
  ]
};

const engine = await CreateWebWorkerMLCEngine(
  new Worker(new URL("./worker.ts", import.meta.url), { type: "module" }),
  "MiMo-V2.6-Distill-Qwen-9B-abliterated-q4f16_1-MLC",
  { appConfig }
);

⚙️ Recommended Generation Parameters

Taken from publisher recommendations and verified on WebLLM:

• Temperature: 0.6 (for reasoning/code), 0.8 - 1.0 (for creative writing)
• Top P: 0.95
• Thinking Mode: Supported! Generates chain-of-thought within <think>...</think>.
• Hardware Requirement: Dedicated GPU with 8 GB+ VRAM recommended (estimated footprint ~6.43 GB with 4K context).