Seven PhD students trained a 7B LLM from scratch in three months and open-sourced everything

Training a large language model from scratch is usually described as a job for teams with serious compute budgets. A project covered by QbitAI offers a different sample: seven PhD students completed a 7B-parameter model in three months, starting from zero.

According to the report, the team published the code, the training data, and the training logs in full, so outside researchers can reproduce the pipeline, inspect how the data was assembled, and check each training decision against the log. That level of transparency is still uncommon in open model releases.

The more discussed detail is how the work was done. QbitAI describes hundreds of agents taking part in the research and engineering process, which is where the “AI starts building AI” framing comes from.

At 7B parameters, the model sits in a range that a small cluster can handle, which makes it a common proving ground for testing training recipes, data mixes, and post-training stages. Seven people closing that loop in three months says something about how much the tooling and automation around training have changed.

The openness matters for a second reason. Releasing code and data lets outsiders trace where a model's abilities come from instead of trusting a leaderboard, and releasing logs turns questions like “which step changed what” into something that can actually be checked.

For the industry, the value of a project like this is less about the model and more about turning the research process into public, inspectable material. Once agents do a large share of the work, who made which decision, and on what evidence, becomes a new axis of evaluation.

What to watch next is whether other teams can reuse this fully documented pipeline, and how far the agent share of research work will move.