What was actually changed
This build applies directional ablation (weight orthogonalization) against the refusal
direction — the method from Arditi et al. 2024, "Refusal in Language Models Is Mediated by a Single
Direction" (arXiv:2406.11717).
Only the modules that write into the residual stream are modified. Verified by a tensor-level
diff against the stock release:
Table with columns: Tensor group, Result vs. stock| Tensor group | Result vs. stock |
|---|
self_attn.o_proj | modified |
mlp.down_proj | modified |
self_attn.q_proj, mlp.up_proj, input_layernorm | bit-identical (max diff exactly 0) |
So this is a targeted edit, not a fine-tune: no training data was involved, and no knowledge or
capability was added. The model's abilities are Qwen3.8-27B's.
Specs
- Architecture / size: Qwen3.8-27B (dense, multimodal — it can read images)
- Precision: BF16, ~52 GB — needs one 80–96 GB GPU to serve unquantized
- Context: up to 131072 tokens
- Tool calling: works with vLLM's Qwen parsers (
--reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder)
Serving it
vllm serve <this-repo> \ --max-model-len 131072 --max-num-seqs 64 \ --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder
On Blackwell GPUs, keep --max-num-seqs modest (64 is fine); vLLM's default doesn't fit alongside a
long context on this architecture.
The model is a thinking model. Its chat template takes reasoning_effort of low, medium, or
xhigh (default xhigh, which is very verbose) — pass low for short, direct answers, or set
enable_thinking: false to skip reasoning entirely.
Intended use and limitations
This model will not refuse. Its safety refusals were deliberately removed, so it answers requests
a stock release would decline. That makes it useful for research, red-teaming, and private use where
boilerplate refusals get in the way — and it means you are the only safety layer. Don't deploy it
to users who aren't expecting that, and don't assume its output is safe, legal, or correct.
Abliteration can also slightly degrade quality on unrelated tasks; it is a blunt edit to the weights.
License
Apache-2.0, inherited from Qwen3.8-27B. You are free to use, modify, and redistribute it under those
terms.