Open weights keep shipping: Inspect, Phonon-2, Kumo Tabular and a volunteer 145M model

The release landed at 15:20 UTC on 30 September, according to the project's own site. Inspect is a framework for frontier AI evaluations, built by the UK AI Security Institute and Meridian Labs. Its selling point is breadth: coding, agentic tasks, reasoning, knowledge, behaviour and multi-modal understanding, all in one tool.

The framework ships composable datasets, agents, tools and scorers, plus a library of over 200 pre-built evaluations. There is a web-based Inspect View tool and a VS Code extension. Sandboxing runs untrusted model code in Docker, Kubernetes, Modal, Proxmox or Vagrant through an extension API. The project says it supports more than 20 model providers, including local inference with HuggingFace, vLLM and SGLang.

Evaluation tooling is the unglamorous half of open weights. Weights without a way to measure them are just files.

A 164 MB speech model and a tabular model that needs no training

Fermion Research released Phonon-2 on 30 September and called it the most accurate open speech recognition model under 900 MB. The download is 164 MB. The company says it averages 5.21 percent word error across the seven English sets on the Open ASR Leaderboard, and that every open model scoring better is at least 5.8 times its size.

The encoder stores each weight as one of five learned levels, roughly 2.1 bits, packed as base-3 digits five to a byte. The weights are under CC-BY-4.0, the licence of Nvidia's Parakeet TDT 0.6B v3, from which they derive. Fermion claims an hour of audio becomes text in about 20 seconds on a MacBook Air, and that an H100 handles 13 seconds of audio per day in batches of 128.

Nvidia's contribution is a different kind of compression. Kumo Tabular, published on Hugging Face on 29 September, predicts labels from a labelled table in a single forward pass with no training, no tuning and no feature engineering, for classification and regression. It comes in three sizes from 28M to 215M parameters, 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.

Volunteers, Rust, and a 145M model trained on cron jobs

Smaller efforts are moving too. Coop, a GitHub project updated on 30 September, is pretraining a roughly 145M-parameter decoder-only transformer on FineWeb-Edu using donated consumer hardware and free tiers of Hugging Face and GitHub Actions. Pseudo-gradients arrive as pull requests against a public dataset repo. A stateless GitHub Actions cron aggregates them with a DiLoCo-style outer step every five minutes in theory, though the project notes GitHub's scheduler fires anywhere from minutes to hours apart. Stage 1, a 15M model on TinyStories, ran past its Chinchilla-optimal budget in six days and took validation loss from 9.01 to 2.8.

Separately, a Rust project called PSSA, also updated on 30 September, describes a non-transformer architecture with a selective state-space recurrence, an episodic memory bank and per-token weight updates. The author claims it learns faster than a transformer at matched parameters and generates about twelve times quicker on the same CPU. There is no independent benchmark in the repository.

The week's open releases sit against a harder regulatory backdrop. The Guardian reported on 30 September that the FTC has opened an industry-wide investigation into Anthropic, OpenAI and others, the first official US enforcement action touching rogue AI agents. Ars Technica reported the same day that OpenAI is delaying its IPO over safety concerns. Open weights are not the subject of either action, but the scrutiny sets the tone for everything shipping alongside them.