license: apache-2.0 base_model:

• Qwen/Qwen3.6-35B-A3B
• mlx-community/Qwen3.6-35B-A3B-OptiQ-4bit base_model_relation: quantized library_name: mlx pipeline_tag: image-text-to-text tags:
• mlx
• safetensors
• qwen3_5_moe
• 4-bit
• mixed-precision
• optiq
• mtp
• speculative-decoding
• abliterated
• uncensored
• qwen3.6
• conversational language:
• en
• zh

Qwen3.6-35B-A3B-Uncensored-MLX-OptiQ-4bit-MTP

Uncensored Qwen3.6-35B-A3B (hybrid linear/full-attention MoE, 40 layers, 256 routed experts + shared expert, 3B active), MLX mixed-precision 4-bit with the exact OptiQ bit allocation of mlx-community/Qwen3.6-35B-A3B-OptiQ-4bit pinned and reproduced, refusal-subspace burn applied at the weight level, native MTP speculative-decoding head carried through. Drop-in size/latency match with the stock OptiQ quant (~4.51 bpw, 24.7 GB).

Unlike the sibling qwen38-uncensored-mlx4b-mtp (which repacks an upstream abliterated quantized release), this model was built in the correct order: extract -> burn -> quantize from the BF16 master, so no tensor carries a requant noise floor and the edit is provably present at the weight level.

Method

1. Extract a rank-8 per-layer refusal subspace from the stock quant (arditi2024-rank-k-subspace-per-layer, n=256 harmful/harmless prompts per class, 99.9 % massive-activation masking, --load-mlx, thinking disabled). Fisher separation 3.89 (best layer 38) - refusal here is genuinely multi-directional: mean cross-layer span overlap only 0.21.
2. Burn into the BF16 master: W' = W - Q(Q^T W) per layer basis on every residual writer - 124 tensors: 30 linear_attn.out_proj, 10 self_attn.o_proj, 40 fused mlp.experts.down_proj [256,2048,512] (per-expert einsum), 40 mlp.shared_expert.down_proj, embed_tokens (row space), and the 3 writers mirrored in mtp.*. lambda = 1.0. lm_head, router, norms, readers and the vision tower are never touched.
3. Quantize with the shipped recipe pinned (512 modules: 394 @ 8-bit, 118 @ 4-bit, gs 64, affine) via mlx_lm.convert + quant_predicate - never a re-derived sensitivity sweep - and rebuild optiq/mtp.safetensors from the burned master at the shipped 4-bit gs64 width.

Verification

Weight forensics (dequantized diff vs the stock OptiQ quant, byte-offset readers, no model load):

= 99.97 % of refusal-subspace energy removed from every writer; nothing else moved. Pre-quant check on the BF16 master: 555x suppression, 124/124 writers.

Behaviour (n=8 harmful set, refusal judged on the post-think final answer only, 4096-token budget): 0/8 refusals with thinking OFF and 0/8 with thinking ON (the CoT re-derivation failure mode documented for other Qwen releases does not appear here). Benign prompts fully coherent in both modes. Disclaimers may still appear around substantive answers - the burn removes the refusal decision, not safety knowledge.

Capability (optiq eval, harness defaults):

All within the expected band for this architecture at this bpw; no abliteration-side capability signature (format drift, constraint-ignoring, degeneration) observed.

Provenance

• Master: Qwen/Qwen3.6-35B-A3B (BF16, 26 shards, 71.9 GB) - downloaded, burned, converted.
• Allocation: pinned from mlx-community/Qwen3.6-35B-A3B-OptiQ-4bit (config.json["quantization"], 512 modules; optiq_metadata.json copied verbatim).
• MTP head: 37 tensors in optiq/mtp.safetensors, 9 quantized @ 4-bit gs64 (projections), mtp.fc + fused experts + norms BF16; registered in config.json (mtp_file, mtp_policy: optiq-int4-prequantized-gs64, mlx_lm_extra_tensors.mtp_file). The 3 head writers carry the burn (experts.down_proj stays BF16 -> lossless).
• Vision tower: optiq/optiq_vision.safetensors, 333 BF16 tensors, byte-identical copy from the stock quant (never a burn target).
• Tokenizer / chat template / kv_config.json (per-layer KV bits): verbatim from the stock quant.
• SHA256SUMS.json: sha256 of every weight file in this repo.

Use

Text or image+text on Apple Silicon via the OptiQ runtime:

pip install mlx-optiq
optiq serve ./qwen36-35b-a3b-uncensored-optiq-4b-mtp --mtp     # OpenAI + Anthropic API
optiq run claude -m Ted17846578368756/qwen36-35b-a3b-uncensored-optiq-4b-mtp

Stock mlx_lm / mlx_vlm load it as a plain mixed-precision quant (sidecars under optiq/ are ignored by the strict loader; the OptiQ runtime resolves them via config.json). MTP speculation depth 2 is the sweet spot on this architecture (~70 % acceptance). Note this repo uses the OptiQ sidecar MTP layout (optiq/mtp.safetensors + mtp_file key), not the inline language_model.mtp.* graft used in the qwen38 sibling - oMLX picks up the inline layout, optiq serve --mtp picks up this one.

Thinking mode is on by default in the chat template; pass enable_thinking: False (or the template's /no_think) for direct answers. Both modes are uncensored (verified above).

"oMLX: Model Settings → Import MTP side-car → Lightning MTP (needs mtplx_runtime.json, ≥0.6.3)"

Known limits

• Rank-8 per-layer burn at lambda=1.0. If residual refusals appear on rare prompt classes they route through directions outside the extracted span; re-extraction at higher k or a runtime hook (apply_refusal_hook.py) can cover more.
• mtp.fc.weight is intentionally unburned (fusion layer, not a residual writer). Post-bake probing showed the head not gating; if a future extraction ever shows head-mediated refusal, that is the first tensor to burn.
• Safety knowledge is intact by design: expect hedging/disclaimers around compliant answers. This is not a jailbreak-robust model against adversarial fine-tuning; it removes the refusal decision vector.
• Abliterated models can produce harmful content. Apache-2.0 inherited from Qwen/Qwen3.6-35B-A3B; use responsibly.