Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
›

Open weights go compact: Phonon-2, Kumo Tabular and a volunteer-trained model ship

Fermion Research released Phonon-2 on 30 September. It is an open speech recognition model that fits in a 164 MB download. It averages 5.21% word error across seven English sets on the Open ASR Leaderboard. Within hours, Nvidia's Kumo Tabular and a volunteer-trained 145M-parameter language model landed alongside it.

AI & modelsNewsGrace OkonkwoPublished: 30 September 20263 min readSources 7
Open weights go compact: Phonon-2, Kumo Tabular and a volunteer-trained model ship

Three open-weight releases landed inside a single day. None of them chased scale. Fermion Research published Phonon-2 on 30 September. It is an English speech recognition model that stores each encoder weight as one of five learned levels at about 2.1 bits, in a 164 MB file. The company says it matches its 2.5 GB full-precision teacher set for set. It beats the teacher on meetings and parliamentary speech. It transcribes an hour of audio in about 20 seconds on a MacBook Air. The weights are on Hugging Face under CC-BY-4.0. That is the licence of Nvidia's Parakeet TDT 0.6B v3, from which they derive.

That last detail matters. The small open-model race is largely a race of teachers and licences.

Fermion claims every open model that scores better on the Open ASR Leaderboard's seven English sets is at least 5.8 times its size. It says the lead holds under noise, staying ahead of Parakeet Redux, the other low-bit model of its size. Speed figures come with conditions attached. They cover one clip at a time, log-mel included, model load excluded. Each fast path ships only if its word error stays within noise of the exact path. That check uses a paired bootstrap over clips with 4,000 resamples.

Kumo Tabular attacks the tree-model lifecycle

Nvidia published Kumo Tabular on 29 September. It is an open foundation model for tabular classification and regression. It predicts new rows in a single forward pass with no training, tuning or feature engineering. It comes in three sizes from 28M to 215M parameters. It is released under the OpenMDW-1.1 licence for commercial use. It ranks first on four benchmarks: TabArena, BeyondArena, TALENT and ScoringBench, according to the Hugging Face post.

The architecture borrows from TabICL and TabPFN, with column, row and in-context attention. The company says it was pretrained only on artificial data. That sidesteps the privacy problem of training on customer tables. Whether it unseats gradient-boosted trees in production is a separate question from benchmark placement. Nvidia's post does not answer it.

A language model trained on GitHub Actions and donated hardware

The Coop project is a GitHub repository that went live on 30 September. It is attempting something stranger: pretraining a small language model with no servers and no funding. Stage 2 is a roughly 145M-parameter decoder-only transformer training from scratch on FineWeb-Edu. Workers download a checkpoint from Hugging Face. They run local AdamW steps on a personal data shard. They submit a pseudo-gradient as a pull request against a public dataset repository. A stateless GitHub Actions cron job runs every five minutes. It aggregates the submissions with a trimmed mean or geometric median and takes one Nesterov outer step.

Stage 1 is the proof. A 15M-parameter model trained past its Chinchilla-optimal budget by volunteers in six days. Validation loss fell from 9.01 to 2.8. The weights stay available. The project's own documentation is unusually candid about evaluation. It notes that a single validation loss number says nothing, because outer steps move it up as often as down. The direction is a property of the series, not one point.

Two other open releases rounded out the window. The UK AI Security Institute and Meridian Labs published Inspect. It is an evaluation framework with over 200 pre-built evaluations. It also offers sandboxing for untrusted model code in Docker, Kubernetes and Modal. And Sparticle62ops released PSSA, a non-transformer language model written in Rust with no ML framework underneath. It claims PSSA learns faster than a transformer at matched parameters. It also claims PSSA generates text about twelve times quicker on the same CPU.

Open weights are not the whole story. Also on 30 September, the FTC opened an industry-wide investigation into Anthropic, OpenAI and the research group Metr. It is the first official US enforcement action on rogue AI agents, as The Guardian reported. OpenAI, meanwhile, is delaying its IPO and holding back a model release over safety concerns, per Ars Technica. The contrast is hard to miss. The models shipping without friction are the small ones. The companies under scrutiny are the ones with the largest valuations.

Comments 0

Sources

7
  1. 01Phonon-2: most accurate open speech recognition model in a 164 MB downloadEN
  2. 02Nvidia Kumo Tabular: Open Foundation Model for Tabular PredictionEN
  3. 03Coop: A small language model pretrained by volunteersEN
  4. 04Inspect: An open-source framework for large language model evaluationsEN
  5. 05PSSA: A non-transformer language model written from scratch in RustEN
  6. 06US trade regulator opens investigation into AI giants including Anthropic and OpenAIEN
  7. 07OpenAI delays IPO over AI safety concernsEN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Grace Okonkwo

Grace Okonkwo

AI, models and technology

Grace Okonkwo covers AI, models and technology for FLASH24, working from primary sources such as model cards, API documentation and benchmark papers rather than vendor summaries. She checks training data provenance, evaluation conditions and reported scores against the underlying datasets before any figure reaches print. She interviews researchers and engineers directly, tracks release calendars from major labs, and compares successive model versions on the same tests. Her own self-hosting, home-network and documentation-reading habits feed straight into that desk, since she tests tools on her own hardware first. She does not publish benchmark claims without a reproducible method.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.