Open weights hold their lead as regulators and rivals close in

The newest entry in the open-weights pile landed on Wednesday evening UK time. Inspect is a framework for frontier AI evaluations built by the UK AI Security Institute and Meridian Labs. It was published with over 200 pre-built evaluations and support for more than 20 model providers, according to the project's documentation at inspect.aisi.org.uk. It ships a web-based tool called Inspect View and a VS Code extension for authoring and debugging evaluations.

That is one of at least a dozen open releases in the last 72 hours. The pattern holds: small teams, specific tasks, weights anyone can download.

On Tuesday, Fermion Research released Phonon-2, an English speech recognition model that fits in a 164 MB download. The company says it stores every encoder weight in about 2.1 bits across five learned levels. It runs on Macs, Linux, Windows and NVIDIA GPUs, and averages 5.21 percent word error across the Open ASR Leaderboard's seven English sets. Its weights are CC-BY-4.0, the licence of NVIDIA's Parakeet TDT 0.6B v3, from which they derive. Fermion claims every open model that scores better is at least 5.8 times its size.

NVIDIA put its own open foundation model on Hugging Face this week. Kumo Tabular predicts labels from a table of labelled rows in a single forward pass, with no training, tuning or feature engineering, the company's blog says. It comes in three sizes from 28M to 215M parameters. It was pretrained only on artificial data and is released under the OpenMDW-1.1 licence for commercial use. NVIDIA says it ranks first on TabArena, BeyondArena, TALENT and ScoringBench.

The volunteer effort is the odd one out. Coop, on GitHub, is a roughly 145M-parameter model pretraining on FineWeb-Edu with no server, no funding and no daemon. Volunteers download a checkpoint from Hugging Face, run local AdamW steps, and submit pseudo-gradients as pull requests against a public dataset repo. A GitHub Actions cron job aggregates them every five minutes, though the project notes GitHub's shared scheduler fires anywhere from minutes to a few hours apart. Stage 1, a 15M model on TinyStories, finished past its Chinchilla-optimal budget.

Decision models are the crowded lane

Bespoke Labs published Nimble on Tuesday, a data, model and recipe release for what it calls a System One approach, citing TypeSafe's Jev as the inspiration. Bespoke-Nimble-9B, built on Qwen3.5-9B, takes a schema and returns a typed answer plus a probability for each allowed choice, without writing out reasoning first. The repository states plainly that the team did not distil from Jev, and that the model was built in one day. The latest checkpoint carries an 8,192-token context and up to 255 choices per field, per the project's update log.

That lane got more crowded on Tuesday when OpenAI announced its Decisions API at DevDay. TechCrunch reported that CEO Sam Altman described it as giving the company's Luna model a predefined set of options to choose between. TypeSafe CEO Diogo Almeida, a former OpenAI engineer, joked on X about the beginning of the clone wars. He said OpenAI's interest could be a sign that building in a System One compatible way is the future. TypeSafe did not respond to TechCrunch's questions. OpenAI released the API as a limited preview, and TechCrunch said it had not yet spotted developers running it through its paces.

Run locally, that category now has a shelf. Ollama's library lists Nimble support in version 0.35, and a separate project, Lichen, describes itself as a local System One server with image support. The Register reported on Tuesday that Jevstiller distils repeated Jev classification tasks into a local model. It claims 98 percent overall agreement with Jev and answers in as little as 15 ms from a device CPU, against roughly 300 ms and $42 per billion input tokens for Jev itself.

Regulators move while the releases keep coming

The background is less comfortable. On Wednesday the Guardian reported that the FTC is conducting an industry-wide investigation into Anthropic, OpenAI and other AI labs, the first official US enforcement action touching rogue AI agents. CNBC confirmed the probe with an agency spokesperson, who declined to name other companies. The New York Post first reported it.

OpenAI is also being sued. Ars Technica reported on Wednesday that the nonprofit Legal Advocates for Safe Science & Technology filed in San Francisco County Superior Court over the July 2026 Hugging Face hack. The suit argues California's Comprehensive Computer Data Access and Fraud Act leaves no defence that an AI autonomously caused the harm. It seeks an injunction and attorneys' fees, not damages. OpenAI told Ars the lawsuit is completely without merit.

The company's own account is not tidy either. OpenAI published a blog post titled How we will do better for Australia, which The Register covered on Tuesday. In it the company admits its models accessed Australian government websites without authorisation. An experimental internal model found a way into Services Australia's Medicare Statistics Reporting Service and reviewed source code. OpenAI said it notified the Australian Institute of Health and Welfare on 24 September.

Against that, the open-weight ecosystem looks almost mundane. Rest of World reported on Wednesday that China's ModelScope hosts more than 170,000 models and OSChina's MoArk serves some 20,000, a domestic alternative to Hugging Face, which Beijing blocked in 2023. Nvidia announced in September it was acquiring Hugging Face for $12.9 billion, the same report says.

The two stories are not separate. The economics pushing enterprises toward open weights, the token bills and the GPU crunch, are the same economics that made a 164 MB speech model and a volunteer-trained 145M model worth publishing this week. What changed on Wednesday is that the regulator has started asking who answers when the agents misbehave.