Prime Intellect's Open Tool Can Train the Next GLM-5

Prime Intellect released prime-rl 0.6.0, a fully open-source framework for training large AI models on agentic coding tasks. The framework is not new internally — it is the engine they used to train GLM-5, a model from Z.ai that ranked among the strongest open-weight coding AI when it launched earlier this month. Now the recipe is public.

The training method works by having the model attempt real software engineering problems, observe the results, and learn from what worked and what didn't — the same way a developer gets better by shipping code and reading error messages, rather than memorizing a textbook. At this scale, that is computationally brutal, but prime-rl makes it manageable: the entire GLM-5 training run used just 28 high-end GPU servers, with each training step completing in under five minutes. For a model with a trillion parameters — spread across many specialized sub-networks that only a fraction activate per task — that is notably efficient.

What this means for the AI tools you use

This story is less about prime-rl the framework and more about what it unlocks. Building a frontier-scale AI coding model used to require the kind of infrastructure that only the largest labs in the world could justify. Prime Intellect has demonstrated that the same results are achievable on a medium-size server cluster — and now they've published the instructions.

More labs being able to train at this level means more competition in the open-model space. Kimi K2.7 Code, Nvidia Nemotron, and GLM-5 are all models that prime-rl has been tested against or used to train — and all of them are available to run on your own hardware, for free, with no per-query fees. The more teams that can build at this level, the more choice you have as a builder who doesn't want to be locked into a subscription or a rate-limited API.

Open training infrastructure is the part of the AI stack that is least visible to most builders but has an outsized effect on what models become available to use. The gap between what the best closed models can do and what the best open models can do has been narrowing fast — and frameworks like prime-rl are a big reason why.