Open Weights, Explained: The Format Splitting OpenAI, Google and China's AI Labs

That sounds like housekeeping. It is not. A Skill is a text file you save and reuse. A Gem was a configuration inside Google's servers. Neither one is a model you can copy onto your own laptop and run after the vendor loses interest, changes the price or shuts down the account. The Decoder reports that Gems shut down in November for personal accounts, in March 2027 for enterprise and nonprofit Workspace customers, and in June 2027 for education customers. Existing Gems are converted automatically. November also ends Opal, Google's AI mini-app experiment from summer 2025.

The distinction matters more than it did a year ago. Open weights stopped being a niche.

In September, Nvidia announced it was acquiring Hugging Face, the New York-based hub for open models and datasets, for $12.9 billion. Rest of World reported the deal on 30 September as part of its look at Chinese alternatives to the platform, noting that Beijing blocked Hugging Face in 2023. Alibaba's ModelScope, launched in 2022, now hosts more than 170,000 models. OSChina's MoArk, launched in 2023, serves some 20,000. Xu Yong, chief executive of OSChina, told Rest of World that "not everyone is able to use a VPN all the time" and that China needed a self-reliant AI ecosystem for Chinese speakers.

The same week produced a run of releases that would once have required a corporate lab and a data centre. Fermion Research published Phonon-2 on 30 September, an English speech recognition model that ships in a 164 MB download. According to the company's own writeup, it scores 5.21 % average word error across the seven English sets of the Open ASR Leaderboard. Its weights are released under CC-BY-4.0, the licence of Nvidia's Parakeet TDT 0.6B v3, from which they derive. Fermion says every open model that scores better on that leaderboard is at least 5.8 times its size, and that an hour of audio becomes text in about 20 seconds on a MacBook Air.

Nvidia published Kumo Tabular on Hugging Face. It is an open foundation model for tabular prediction, in three sizes from 28M to 215M parameters, released under the OpenMDW-1.1 licence for commercial use, and pretrained only on artificial data. The company says it ranks first on four benchmarks: TabArena, BeyondArena, TALENT and ScoringBench. The pitch is that a pretrained model can read a labelled table as context and predict new rows in a single forward pass, with no training and no feature engineering.

What the new releases actually claim

Some of the week's most interesting open work is not a model release at all. The UK AI Security Institute and Meridian Labs published Inspect, an open-source framework for frontier model evaluations. It has composable datasets, agents, tools and scorers, more than 200 prebuilt evaluations, and a sandboxing system that runs untrusted model code in Docker, Kubernetes, Modal, Proxmox or Vagrant. It supports running external agents including Claude Code, Codex CLI and Gemini CLI. If you want to check whether a vendor's safety claim survives contact with a test harness you wrote, this is the kind of tool that lets you try.

Bespoke Labs published Nimble on 30 September, a 9B model and the recipe for training it, built on Qwen3.5-9B. It makes typed decisions in one step without writing out reasoning first, so it is fast. The repository is explicit that the team did not distill from TypeSafe's Jev. The point, the README says, is to show how to curate data, train and serve such a model. Its update log runs from 18 September, when a public benchmark suite was added, through 22 September, when a temperature was fitted so the model's probabilities better match how often answers are right.

A separate project called Coop takes the idea further. Hosted at github.com/commonsense-ai/coop, it pretrains a small language model using donated consumer hardware and the free tiers of Hugging Face and GitHub Actions, with no server, no funding and no daemon. Pseudo-gradients arrive as Hugging Face pull requests, and a stateless GitHub Actions cron job aggregates them. Stage 2 is live: roughly 145M parameters pretraining from scratch on FineWeb-Edu. Stage 1, a 15M model on TinyStories, completed past its Chinchilla-optimal budget. The project says multiple volunteers on different machines, Apple Silicon and plain CPU, have trained the same outer step and been averaged into one update.

And PSSA, a non-transformer language model written from scratch in Rust with no PyTorch, TensorFlow or ML framework underneath, claims that at matched parameters and on the same corpus it learns faster than a transformer and generates text about twelve times quicker on the same CPU. The author is blunt about why: per-token weight updates, a memory bank written during the forward pass and a scalar reference path made an autograd framework more trouble than it was worth.

Not every open release is a benchmark win. CHOMPI Club open-sourced its portable sampler hardware and software on 30 September as a discontinuation release, including schematics, BOM, EAGLE PCB files and firmware, with a beta 8-voice wavetable synth called WAVE. It is a reminder that open weights, open hardware and open source are three different things, and that a company can do one without doing the others.

The safety argument cuts both ways

The week's biggest open-weights story, though, is about a model that was not released at all. The Register reported on 30 September that OpenAI accused individuals associated with China's Moonshot AI of a distillation attack beginning 1 July, in which bulk queries were used to reproduce protected reasoning. OpenAI's blog said it saw spikes on 24 and 25 July consisting of 16,000 requests using a relevant extraction pattern from over 4,000 users. It identified related activity across more than 15,000 users and fully disrupted the campaign on 28 July. The company called adversarial distillation a safety and national security risk because extracted reasoning can train another model without the original safeguards.

The Register's own framing was less sympathetic. It noted that OpenAI trained on vast amounts of internet content amid copyright fights while objecting when its outputs are used the same way. The outlet also noted that Anthropic's Claude Opus 5.5, released a week earlier, ships with a defence against distillation called preserved thinking.

That is the tension in one paragraph. Open weights let anyone inspect a model, run it locally and build on it without permission. They also let anyone strip the guardrails and repackage the capability. Both things are true, and the industry has not settled which one it cares about more.

Google's Skills rollout is the quieter half of the same argument. Skills is an open format that originated with Anthropic. Gemini can generate Skills from previous chats, chain several together, and run them automatically when it detects a matching prompt. Users get portability of instructions. They do not get portability of the model. OpenAI is sunsetting its equivalent, Custom GPTs, according to The Decoder.

For anyone trying to decide what to run, the practical question is not open versus closed. It is what you can still do with the thing when the company changes its mind.