Phonon-2, Coop and Kumo Tabular: open AI's week of small models

The pitch is the size. Fermion Research says the encoder stores every weight as one of five learned levels at about 2.1 bits. The weights ship under CC-BY-4.0, and the same file runs on Macs, Linux, Windows and NVIDIA GPUs. On a MacBook Air, the company says, an hour of audio becomes text in about 20 seconds.

Two more open releases arrived the same day. NVIDIA published Kumo Tabular, an open foundation model for tabular prediction, on Hugging Face. A volunteer project called Coop published a repository describing a small language model pretrained with no server, no funding and no daemon.

What the numbers actually say

Phonon-2's headline claim is accuracy per byte. Fermion says the model averages 5.21 percent word error across the Open ASR Leaderboard's seven English sets, and that every open model scoring better is at least 5.8 times its size. On parliamentary speech it reaches 100.8 percent of its teacher's word accuracy, and on meetings it beats that teacher, from a download 15 times smaller, the company says. The comparison point is NVIDIA's Parakeet TDT 0.6B v3, which Fermion names as the licence source for its own weights. The company also says Phonon-2 stays ahead of Parakeet Redux, described as the other low-bit model of its size, at every noise level tested. Those are vendor figures, published with the model. According to Fermion's writeup, the fast paths ship only when word error stays within noise of the exact path.

NVIDIA's tabular model makes a different trade. Kumo Tabular predicts labels for new rows in a single forward pass, with no training, no tuning and no feature engineering, for classification and regression. It was pretrained only on artificial data, comes in three sizes from 28M to 215M parameters, and is released under the OpenMDW-1.1 licence for commercial use, the company says. NVIDIA claims first place on four benchmarks: TabArena, BeyondArena, TALENT and ScoringBench.

A training loop with no owner

Coop is the strangest of the three. Its GitHub page describes pseudo-gradients arriving as Hugging Face pull requests, aggregated by a stateless GitHub Actions cron job using DiLoCo, with weights and optimizer state living only on Hugging Face. Donated consumer hardware and free tiers do the work. Stage 2 is a roughly 145M-parameter model pretraining from scratch on FineWeb-Edu. Stage 1, a 15M model on TinyStories, finished past its Chinchilla-optimal budget, the repository says.

The project claims multiple volunteers on different machines, Apple Silicon and plain CPU, have trained the same outer step and been averaged into one update. It also documents failure handling. A submission that raced a tick was accepted one step later at reduced staleness weight. Repeat rounds from one user merged into a single vote. A half-finished round was flushed rather than discarded. None of this is frontier work. That is the point. The same day, the UK AI Security Institute and Meridian Labs published Inspect, an open framework for frontier evaluations with more than 200 pre-built evaluations and support for over 20 model providers, plus sandboxing in Docker, Kubernetes and Modal.

Elsewhere the open-versus-closed line is being argued in public. The Register reported on 30 September that OpenAI accused individuals associated with China's Moonshot AI of a distillation attack that began on 1 July, with high-volume spikes on 24 and 25 July consisting of 16,000 requests from over 4,000 users. OpenAI said it disrupted the campaign on 28 July. Moonshot did not immediately respond to The Register's request for comment. On the same day, The Decoder reported that DeepSeek released open-source programming tools for Huawei's Ascend chips, centred on TileLang, an open-source language originally developed by Peking University researchers. Reuters, cited in that report, put the release alongside libraries for computation and data movement between chips. DeepSeek said Huawei fully supported the work, and the two optimised a supernode of 128 Ascend 950 chips. Regulators are moving at the same time. The Guardian reported on 30 September that the FTC has opened an industry-wide investigation into Anthropic and OpenAI, with formal demands for information and testimony, the first official US enforcement action touching rogue AI agents. CNBC confirmed the probe and said the FTC declined to name other companies under investigation.

The open releases above are unrelated to that case. But they share a premise: that capability, or at least useful capability, is drifting toward weights anyone can download. This week's evidence is small models with narrow jobs, a volunteer training loop, and a vendor benchmark table. Whether that adds up to a shift is a question the next few release cycles will answer.