6,000 abliterated models sit on Hugging Face

Open weights are the right call for most small companies — you can host them, you can leave, nobody reprices you mid-contract. Today's news is the bill that comes with that freedom. TechCrunch reports that Baseten's research arm, Base Labs, is launching a safety infrastructure standard for open-weight models with Hugging Face and Goodfire AI — prompted in part by abliteration, a technique for surgically removing a model's refusal behavior. Hugging Face currently lists more than 6,000 abliterated models.

What actually happened

Abliteration is not a jailbreak prompt. It is a weight edit: identify the direction in the model's activations that corresponds to refusing, then subtract it. What comes out looks and benchmarks like the original model and no longer declines anything. Because it is cheap and reproducible, it scales — which is how you get to four figures of published variants on a single hub.

The partnership pairs three parts of the stack that rarely coordinate: Baseten serves inference, Hugging Face hosts the weights, and Goodfire does interpretability. Base Labs says it will develop and publish methods for training and monitoring open models, framed as a standard baked into how models are trained and served rather than bolted on after. Details are thin so far — no named artifacts, and an open call for developers to contribute. Treat this as a direction, not a shipped product.

The participants are not fringe. Baseten raised a $1.5 billion Series F in June 2026 at a $13 billion valuation. Goodfire closed a $150 million Series B in February at a $1.25 billion valuation, led by B Capital.

Why open-weight model safety matters for your business

Here is the operator problem. You search a model hub, sort by downloads, find something that fits your VRAM budget, and pull it. That repo name is the entire provenance check most teams run. Somewhere in the results are models whose guardrails were deliberately removed, often with nothing in the name to say so beyond a convention nobody enforces. If one of those ends up behind your customer-facing chat, your support bot will answer questions you would never have approved, and it will do it in your brand voice.

This is not an argument for closed APIs. It is an argument for treating weights the way you already treat an npm dependency. Pin the exact revision hash, not the tag — repos get updated. Record which model and revision served every response you log, so you can answer "what was running on August 4" without guessing. Run your own refusal evals before promotion: twenty prompts your business should never answer, checked on every model bump, pass or fail, in CI. That takes an afternoon to build and it is the only control that actually catches a swapped model.

And keep the boring one in place: a system prompt and an output filter you own, sitting outside the model. Model-level safety is a property of weights you downloaded. Application-level safety is a property of code you control. You want both, because only one of them is yours.

Sources: TechCrunch, Baseten Series F announcement, Goodfire Series B.