Coop's volunteer-trained 145M model, and the small-model push around it

The project is called Coop. Its repository describes a training loop that runs entirely on volunteer machines plus the free tiers of Hugging Face and GitHub Actions.

Stage 2 is live: a roughly 145M-parameter model pretraining on FineWeb-Edu. Stage 1, a 15M-parameter run on TinyStories, finished past its Chinchilla-optimal token budget. Volunteers download a checkpoint, run a number of local AdamW steps on a personal data shard and submit a pseudo-gradient as a pull request against a public Hugging Face dataset repo. A GitHub Actions cron job, scheduled every five minutes, reads the open pull requests, drops stale ones, clips and gates the rest, aggregates them with a trimmed mean or geometric median, takes one Nesterov step and uploads a new checkpoint. The repo says GitHub's shared scheduler actually fires anywhere from minutes to a few hours apart, and that the protocol tolerates any cadence.

The project claims its loop is production-proven rather than designed. Multiple volunteers, on Apple Silicon and plain CPU, have trained the same outer step and had it averaged into one update. 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, and a half-finished round was flushed instead of discarded.

That matters because the small end of the market keeps getting more crowded. On 2 October, Microsoft AI released MAI-Transcribe-2-Streaming, a real-time transcription model covering 60 languages that returns first partial results in just over 100 milliseconds, The Decoder reported. The same release included MAI-Voice-2.1, which speaks 23 languages in one voice, and a Flash variant at 150 milliseconds latency and $15 per million characters, down from $22.

Decision models are the other flank. AWS released Strands Decider 2B, an open source model that returns calibrated choices instead of text, built on the torso of Qwen3.5-2B, TechCrunch reported on 1 October. Amazon distinguished engineer Marc Brooker told TechCrunch the class of models makes "a perfect decider for a workflow step." TypeSafe, which named its Jev model after economist William Stanley Jevons, is working on future versions; CEO Diogo Almeida told TechCrunch he did not yet see real competition emerging.

Hardware is the constraint. Apple now ships a Foundation Model on Macs with an M1 chip or newer, reachable from the terminal with the command fm once Apple Intelligence is set up, t3n reported on 2 October. The model answers locally and does not carry context between prompts.

Research is probing where these systems break. A paper submitted on 29 September traces retry loops in diffusion language model agents to masked decoding and proposes a training-free reverse-scoring rule called Reflect Reverse, according to the arXiv abstract. Another paper, submitted on 1 October, reports that model rankings reverse across agent harnesses: on Terminal-Bench 4, Claude leads GPT by 7.94 points in OpenHands but trails by 30.16 points in PI, across 66 evaluated configurations, the authors write.

Chess is the new cheap testbed for prompt optimization, with 1,118 Lichess puzzles and a total study cost of around $800, per a paper submitted on 30 September. The strongest evaluated model, Gemini 3.5 Flash, used as the meta-model, solves only about 55% of puzzles. Elsewhere, a rank-based memory controller cut prompt tokens by 5.9% and lowered bottom-half tail prediction error by 13.6% against a graph-memory baseline on Synthetic Graph World, with a 24.48% token reduction on LongMemEval at accuracy parity, according to its authors.

Coop sits at the far end of that curve: no datacentre, no budget line, a leaderboard fitted over validation loss against tokens rather than outer steps. The repo notes a single validation loss says nothing, because outer steps move it up as often as down. The direction is a property of the series.