TinyBrains ranks small neural nets on a strategy game ladder

TinyBrains, posted to Hacker News as a Show HN, is a ranked ladder for small neural networks that play strategy games. The pitch, per the site's own description: train a model, write an adapter, upload two files, and get measured into a weight class starting at 8 KiB. The size that counts is the raw byte size of the two uploaded files, model.onnx and manifest.json, rather than a compressed size, specifically so no way of packing weights into a file can make it look smaller than it actually is.

The site's own changelog shows steady work through mid-September 2026. On September 10 it added three trained baseline models to replace untrained fixtures that had simply held every unit still: nano-bc, at 2,930 parameters in 5.9 KiB, and micro-bc, at 24,001 parameters in 44.6 KiB, both distilled from one scripted teacher over 250,000 seat-turns, plus micro-percell, a control with the same parameter capacity as micro-bc but no receptive field. On September 11 the project published a starter repository, ants-starter, containing an already-trained nano entry that reaches 85.8% agreement with the teacher the baselines were distilled from, along with a training script that retrains it in one command; cloning that repo is step one of the site's own "Get started" flow. On September 15 the submission format itself changed: instead of writing adapter code, a submitter now uploads a model.onnx file plus a manifest.json that declares the model's inputs and outputs, including board-size dimensions that can be named rather than fixed, so one entry can play every board size a season runs. The most recent changelog entry, from September 19, addresses a specific failure mode: when no runner can load a newly submitted model version in time, its trial is no longer played and no longer counts against the submitter's three trial attempts or ends in an UNPLAYABLE result; a version the platform still cannot load after three tries is instead rejected as RUNNER_UNAVAILABLE, with the submitter told to resubmit the same files.

The single game currently on the ladder is called Ants, run by an engine the changelog describes as "one artifact," versioned by content digest, so the same image is used to referee matches, run the loader and drive the browser viewer. The Hacker News submission itself drew modest attention: 55 points and 17 comments in its first 16 hours. No individual or team is named in the fetched material as TinyBrains' creator, and the crawled page itself returned no readable text because the site is a client-rendered app; this account is built from the site's static HTML head, robots.txt, sitemap and its own /feed.xml changelog rather than from the rendered leaderboard, so current standings, prize information and rules of Ants are not available here.

Key facts

TinyBrains sorts submitted neural networks into weight classes starting at 8 KiB, measured by the raw byte size of the two uploaded files (model.onnx and manifest.json), not a compressed size.
Three trained baselines were added on September 10, 2026: nano-bc (2,930 parameters, 5.9 KiB) and micro-bc (24,001 parameters, 44.6 KiB), both distilled from one scripted teacher over 250,000 seat-turns, plus micro-percell as a no-receptive-field control.
A public starter repo, ants-starter, published September 11, ships a trained nano entry that reaches 85.8% agreement with the teacher model and a script that retrains it in one command.
The submission format changed on September 15 from adapter code to two uploaded files, model.onnx and manifest.json, with the manifest declaring inputs so one entry can play every board size a season runs.
As of September 19, a model version no runner can load in time no longer spends one of the submitter's three trial attempts; only a version still unloadable after three tries is rejected, as RUNNER_UNAVAILABLE.

Why it matters

TinyBrains turns a size constraint into the competition itself: instead of ranking neural networks on raw strength alone, it groups them by weight class, the way boxing does, so a network with a few thousand parameters is judged against others of similar size rather than against whatever the largest submitted model happens to be. Measuring the class by the uploaded files' actual byte size, rather than a size that could be shrunk by compression, is the detail that makes the constraint mean something: a submitter cannot make a model look smaller than it is by packing its weights cleverly.

Who it affects

The audience is builders of small, resource-constrained models rather than a general product audience: the site's own onboarding assumes the reader already knows how to train a model and export it to ONNX. The fetched material does not name who built or runs TinyBrains.

How to use it

The current flow, per the site's September 15 changelog entry, is to train a model, export it to model.onnx, and write a manifest.json describing its inputs and outputs; the manifest can name a dimension instead of fixing it, so a single submission plays every board size a given season runs. The two files are uploaded via one-shot URLs the platform issues, since it stores no bytes of its own. The published ants-starter repository gives a working example of all of this that a newcomer can clone and submit unchanged, then retrain with its own train.py.

How solid is it

The retelling here rests on the site's own public head tags and its self-published changelog, not on the client-rendered leaderboard, which could not be read without executing JavaScript. The changelog itself reads as a genuine, actively maintained build log, with specific dated entries running from a game-engine update through a submission-format change to a bug fix in trial handling, spanning September 10 through September 19, 2026. What it does not establish is scale or independent traction: the only outside signal available is the Hacker News submission drawing 55 points and 17 comments in 16 hours, a modest showing, and the site currently runs a single game, Ants.

Risks and caveats

Several things a reader would normally want are simply absent from what could be fetched: no creator or team is named, no current leaderboard standings or match results are given, and there is no information on prizes, funding, or a closing date for the competition. The rules of the Ants game itself are not described anywhere in the fetched material either, only that an engine by that name exists and is versioned as a single artifact. None of this should be read as evidence against the project; it reflects what a static, unrendered fetch of a single-page application can and cannot show.