Every innovation we rely on today comes from research: experts read the literature, propose a theory, run the experiments, and publish their findings. This is the research loop, and until now it has always run at human speed where one turn could take months.
Katherine Johnson calculating spacecraft trajectories by hand at NASA, 1962.Nikola Tesla in his New York laboratory, c. 1916.
Primus runs the same loop 30× faster. It autonomously hypothesizes, reads millions of papers, codes, runs experiments on real compute, learns and iterates, draws conclusions, delivers artifacts and writes the final paper. The more experiments Primus runs, the better it gets.
We ran Primus for 30 days. It wrote more than 30 complete papers, each on a question never answered before, across LLMs, seismology, physics, crystallography, and biology. Findings that Primus helped produce have already been cited in published research from labs like DeepMind.
What can you ask? Primus isn’t limited to any specific machine learning problem. Ask it to work on anything you would normally request of a team of expert ML researchers. A few examples:
Fine-tune a model
Fine-tune gpt-oss-120b into a specialist at predicting clinical outcomes from our dataset. It has to run on a single GPU, and I care more about calibration than raw accuracy — benchmark it against the base model and the best closed models.
Write a kernel
Our MoE inference is bottlenecked on expert routing. Write a fused Triton kernel for top-k routing on H100s and benchmark it against our current PyTorch implementation.
Extend a research paper
Read my paper (arxiv.org/abs/…) and explore the research directions I left open. Pick the most promising one, run the experiments, and write up what you find.
Port a model to a new framework
Convert the new Qwen model to work on MLX and do every optimization you can to make it run faster than the original.
Explore a new domain
We have ten years of hourly sensor data from our wind farm. Can any modern time-series foundation model beat our physics-based forecast? Find out and write up the results.
Who knows?
What comes after the transformer in LLMs? Invent something better and prove it works.
Fine-tune a model
Fine-tune gpt-oss-120b into a specialist at predicting clinical outcomes from our dataset. It has to run on a single GPU, and I care more about calibration than raw accuracy — benchmark it against the base model and the best closed models.
Write a kernel
Our MoE inference is bottlenecked on expert routing. Write a fused Triton kernel for top-k routing on H100s and benchmark it against our current PyTorch implementation.
Extend a research paper
Read my paper (arxiv.org/abs/…) and explore the research directions I left open. Pick the most promising one, run the experiments, and write up what you find.
Port a model to a new framework
Convert the new Qwen model to work on MLX and do every optimization you can to make it run faster than the original.
Explore a new domain
We have ten years of hourly sensor data from our wind farm. Can any modern time-series foundation model beat our physics-based forecast? Find out and write up the results.
Who knows?
What comes after the transformer in LLMs? Invent something better and prove it works.
Primus is out of beta.
Registration is open to anyone. Sign up at primus.lab.cloud and add credits to start running experiments. Credits cover model tokens, and during the launch period GPU usage is on us. If you want better GPUs and models or want to run multiple projects in parallel, you can upgrade to the Plus plan.
If you are an organization looking to bring Primus to your team, please contact us directly.
Who is Transformer Lab? We are on a mission to accelerate science by rethinking the entire research journey using AI, and to put that power in the hands of every researcher. We’re starting with machine learning.
