AI Pulse · Under the Radar

A 27-billion-parameter model called Qwen3.6-27B-Fable-Fusion just hit 552,026 downloads on Hugging Face with almost no press coverage. It’s built on Alibaba’s Qwen base, then modified using something called “Heretic abliteration” to remove all safety refusals. The result is an “uncensored” model that runs at 4-bit quantization, small enough for consumer hardware, while claiming reasoning scores that usually only show up in proprietary frontier labs.

The model card lists ARC-AGI-Complete benchmarks in ranges where OpenAI and Anthropic typically operate. That’s unverified, but the download count is real. It’s the second most downloaded model in this week’s trending list, right after a Baidu OCR tool. No venture funding, no launch event, no coordinated rollout. Just a model that does what people apparently want: strong reasoning without guardrails, running locally.

For business owners, this matters because the gap between what you can buy from a major lab and what you can download and run yourself is closing faster than most people realize. If a 27B model can genuinely perform near-frontier reasoning on hardware you already own, the cost and control equation shifts. You’re not locked into API pricing or usage policies. You’re also not protected by someone else’s safety layer, which cuts both ways.

This is exactly the kind of shift we track inside the Omni Command Centre, where you can test open-weight models against closed ones and decide what actually works for your use case. The hype cycle moves slower than the models do. By the time something like this gets written up in the trades, it’s already running in production somewhere.