Open weights keep landing: a 164 MB speech model, a 145M volunteer run and China's CUDA alternative

Nvidia put Kumo Tabular on Hugging Face on 29 September. The post is dated 29 September and went to the blog on 30 September. It says the model predicts labels for new rows in a single forward pass, with no training, no tuning and no feature engineering. It comes in three sizes from 28M to 215M parameters, was pretrained only on artificial data, and ships under the OpenMDW-1.1 licence for commercial use. The company says it ranks first on TabArena, BeyondArena, TALENT and ScoringBench.

That is a narrow claim, and Nvidia frames it narrowly: tabular classification and regression, the tables that hold customer records, transactions and sensor logs. The pitch is in-context learning applied to rows rather than text. A model pretrained on millions of tables reads a labelled table as context and returns class probabilities, or 999 quantiles for regression. From those come a point prediction and an uncertainty estimate.

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 encoder stores each weight in about 2.1 bits, and the weights are CC-BY-4.0, the licence of Nvidia's Parakeet TDT 0.6B v3, from which they derive. Fermion reports 5.21 % average word error across the Open ASR Leaderboard's seven English sets, and says every open model scoring better is at least 5.8 times its size. On a MacBook Air it turns an hour of audio into text in about 20 seconds. Those are vendor numbers from a vendor page, so treat the leaderboard position as claimed rather than independently confirmed.

Volunteers, and a stateless cron job

Also on 30 September, a GitHub project called Coop went live with a stage-2 run: a roughly 145M-parameter decoder-only transformer pretraining from scratch on FineWeb-Edu, run entirely on donated consumer hardware plus the free tiers of Hugging Face and GitHub Actions. There is no server and no daemon. Workers download a checkpoint, run local AdamW steps on a personal data shard, and submit a pseudo-gradient as a pull request against a public dataset repo.

A stateless GitHub Actions cron job aggregates each tick. It drops over-stale submissions, clips and cosine-gates the rest, aggregates them, takes one Nesterov outer step and uploads a new checkpoint. The repo claims a 15M-parameter stage-1 model was pretrained past its Chinchilla-optimal budget on TinyStories in six days, with validation loss falling from 9.01 to 2.8. That is a proof of mechanism, not a competitive model.

China's push is the bigger strategic story. Deepseek released open-source programming tools for Huawei's Ascend chips, according to a post on its WeChat channel reported by The Decoder on 30 September and by Reuters. The centrepiece is TileLang, an open-source language originally developed at Peking University, which Deepseek says offers a simpler programming model than CUDA. Huawei "fully supported" the work, Deepseek said, and the two optimised a supernode of 128 Ascend 950 chips.

The context is Nvidia's developer moat: an estimated four million CUDA developers. The Decoder also cites SemiAnalysis, which tested OpenAI's Jalapeno inference chip and called the CUDA moat "potentially dead". SemiAnalysis cautioned that it tested only relatively easy scenarios of about 8,000 input tokens and 1,000 output tokens. Huawei's chips were not part of that comparison.

Regulators move while the weights ship

None of this is happening in a vacuum. The Guardian reported on 30 September that the FTC is running an industry-wide investigation into Anthropic, OpenAI and other labs, with plans to compel testimony from executives including at Metr. The New York Post first reported it. OpenAI, meanwhile, is holding back its own listing. Ars Technica reported on 30 September that Sam Altman will not go public until the company can "make confident safety decisions", and that LASST filed a lawsuit in California the same day.

The open-weight releases above are small, cheap and specific. That is the pattern worth watching, not the benchmark tables.