Community build strips refusals from Qwen3.5-9B and adds DeepSeek-V4-Pro reasoning

A tech blog report on Nov. 7 describes the arrival of an uncensored build of Qwen3.5-9B that reportedly carries DeepSeek-V4-Pro reasoning. The most concrete number in the report is its refusal rate: the model declined only 3 of 100 test prompts.

The naming signals that this is not an official release but a community modification of Qwen3.5's 9B base. In open-weight circles, uncensored usually means safety alignment has been weakened through abliteration, preference fine-tuning or model merging, so that the model keeps answering on topics that would normally trigger a refusal.

The DeepSeek-V4-Pro reference points to the second ingredient: reasoning. Transplanting a stronger model's reasoning style, or its reasoning traces, into a 9B-class model has become one of the most common moves in the small-model community over the past year, trading a larger model's memory footprint for chain-of-thought behavior at a fraction of the cost.

The 3-in-100 refusal rate deserves separate attention because it is both a capability metric and a risk metric. For researchers it means less predictable behavior on sensitive topics; for anyone deploying the weights in a public product, it almost certainly means adding input and output filtering plus a compliance review layer on top.

The distribution pattern is familiar too: an individual blog publishes the evaluation, the open-weight community repeats it, and the story travels faster than any official model card. The cost is uneven granularity, since parameter counts, licenses, data provenance and evaluation methods usually have to be checked against the original weights page.

For people tracking the open-source ecosystem, the thing to watch is not another uncensored model but the 9B tier becoming the main battleground between safety alignment and community modification. It is small enough to run on a single GPU and strong enough to do real work. What follows: whether the Qwen team responds, whether hosting platforms restrict such weights, and whether downstream apps package them into products.