I started with a small question: can I train a language model with RL on my RTX 4060 laptop without running out of memory? That turned into a Prime-RL fork, a separate kernel fork, and an experiment ledger that reaches Experiment 240. Most of the implementation was written with coding agents, mainly Codex, with Claude Code contributing kernel work. I chose the direction, ran into the laptop’s limits, and kept asking whether a faster component actually made training faster.

Three different questions

Those are separate results. A working trainer is useful, but a decreasing SFT loss or a fast decode kernel does not establish a better policy.

Read the documentary

Start with the laptop

The memory budget, the two live policy copies and the guard that keeps the machine usable.

Follow one update

Task, rollout, credit assignment, optimizer step and policy refresh.

Read the experiments

Accepted changes, failed directions and the measurements that could move a graph.

Inspect the evidence

Source links, numeric records, reproducibility limits and the work that remains.
The chapters also cover weights, megakernels, advantages, whole-RL wall time, learning and the installed command. The code map and evidence index connect these explanations to the files and measurements.

Scope of the record

The experiment records here describe October and early November 2026 runs on one laptop. I checked this guide against the available source on 14 November 2026. It is still a work in progress, with no claimed 1B full-training fit or verified long-context coding-agent training result. Source snapshot · Experiment ledger