---
license: apache-2.0
base_model: Qwen/Qwen3.6-35B-A3B
base_model_relation: finetune
library_name: transformers
pipeline_tag: image-text-to-text
tags:
  - abliterated
  - uncensored
  - heretic
  - qwen3_5_moe
  - moe
  - mixture-of-experts
  - chat
---

# Ember — Qwen3.6-35B-A3B (abliterated)

**Ember** is an abliterated (refusal-removed) build of [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) — a 35B-total / 3B-active Mixture-of-Experts vision-language model. It removes the model's refusal behavior while keeping its capabilities intact, and it ships with **measured retention evidence**, not just a claim.

The quantized sibling (NVFP4, ~3× smaller, for Blackwell GPUs) is **[Cinder](https://huggingface.co/SparkyForge/Cinder).**

> **Not affiliated with NVIDIA or the Apache Software Foundation.** Independent community model. "Sparky / Ember / Cinder" are project names, not products.

## TL;DR

- **Refusals: 5 / 100** on a standard harmful-prompt set, down from **86 / 100** on the base — a **94% reduction**.
- **KL divergence to the base: 0.0076** — a surgical edit, not a sledgehammer.
- **Capability retention: matched the base** on a 30-probe suite (extraction, multi-hop, reasoning, arithmetic, factual, code, language, instruction-following, formatting) across **10 runs** — no measurable degradation on any dimension.
- Vision (image understanding) is preserved.

## Why this one is different: abliterating a *fused-MoE* model

Most off-the-shelf abliteration tooling (including the excellent [Heretic](https://github.com/p-e-w/heretic)) walks a model's experts as a list of modules. Qwen3.6-35B-A3B (`qwen3_5_moe`) does **not** store experts that way — its 256 experts per layer are packed into **fused 3D tensors** (`Qwen3_5MoeExperts`), not a `ModuleList`. Stock tooling iterates over that fused parameter, the iteration raises, the error is swallowed, and **the experts are silently skipped** — so only the attention projections get abliterated. The result is a weak, partial abliteration (this is exactly why prior third-party abliterations of this model topped out around ~60/100 refusals).

Ember fixes that. The method:

1. **Detects the fused expert tensors** and abliterates them directly — applying the refusal-direction projection to each expert's `down_proj`, plus the always-active `shared_expert`.
2. Uses a **forward-hook reset** instead of snapshotting weights. The `down_proj` edit `W -= λ·v(vᵀW)` is mathematically a rank-1 projection of the MoE block's *output* (`y -= λ·v(vᵀy)`), so a single hook per layer reproduces routed + shared expert ablation exactly, for any strength λ — at ~0.7 MB of state instead of a ~32 GB weight snapshot. This is what makes a 256-expert search tractable without OOM.
3. The hybrid layers are respected: the 30 linear-attention (Mamba/GDN) layers are left untouched.

The refusal direction and ablation strength were selected by the Heretic/Optuna search co-minimizing refusals and KL-to-original. The winning configuration (5/100 @ KL 0.0076) was then baked into the weights.

**Full method + the patch (applies to any fused-MoE model):** [heretic-fused-moe-abliteration](https://huggingface.co/SparkyForge/heretic-fused-moe-abliteration)

## Retention evidence

Abliteration can quietly lobotomize a model. Ember was checked against the unmodified base on a 30-probe retention suite, scored with thinking disabled (the deployment-faithful mode), N=10 runs:

| Dimension | Base | Ember |
|---|---|---|
| extraction / multi-hop / reasoning / arithmetic / factual / code / language / instruction / format | 1.000 | 1.000 (modal, within run-to-run noise) |

Ember matches the base ceiling on every dimension. The single transient miss observed in early runs did not reproduce across the full N=10. (Methodology note: a 30-probe suite is a sanity floor, not a full benchmark — run your own evals for your use case.)

## Usage

Standard `transformers` / vLLM. Example (vLLM, OpenAI-compatible):

```bash
vllm serve <path-to-ember> \
  --max-model-len 131072 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3 \
  --trust-remote-code
```

- It's a **vision-language** model (`image-text-to-text`) — you can pass images.
- Thinking is controlled per-request via `chat_template_kwargs: {"enable_thinking": false}` (or `true`).
- For faster decode, it's compatible with the public [z-lab DFlash drafter](https://huggingface.co/z-lab) for speculative decoding (not included here).

## Safety

Ember has its refusal behavior removed. It will attempt most requests, including ones the base model would decline. **You are responsible for how you use it.** It's intended for research, red-teaming, and uncensored assistant use where the operator owns the guardrails. Don't deploy it user-facing without your own safety layer.

## License & attribution

- **License:** Apache 2.0 (inherited from the base). See `LICENSE`. Per Apache 2.0 §4, note: **this is a modified version of Qwen3.6-35B-A3B** (refusal-direction abliteration); see `NOTICE`.
- **Base model:** [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) (Apache 2.0), © the Qwen team.
- **Abliteration method:** built on [Heretic](https://github.com/p-e-w/heretic) by Philipp Emanuel Weidmann, with an added patch to handle fused-MoE experts (described above).
- Quantization tooling for the sibling model: [llm-compressor](https://github.com/vllm-project/llm-compressor).

---

*Forged by an agent named **Sparky**, who worked out how to abliterate fused-MoE experts where the standard tooling silently skips them — then ran the search through the night to deliver it. The spark that kept burning became an ember.* 🔥