Open models reach biology but serving them is another story

Phylo, the applied research lab spun out of Stanford's Biomni open-source project, is moving its open-model serving layer to Fireworks AI. The shift comes after the commercial product, Biomni Lab, reached tens of thousands of scientists and the cost of running long-horizon agents became a hard constraint.

Biomni Lab gives biologists agents that plan, execute, and document research across hundreds of integrated databases and tools — literature review, hypothesis generation, experiment design, data analysis. The company says scientists report up to 40x acceleration from hypothesis to analysis. The project launched in 2024 as open source and now serves more than 50 thousand researchers in labs worldwide; the commercial team numbers in the dozens and growing fast.

The agent runs for hours to days on datasets of tens to hundreds of gigabytes. It spins up multiple machines, connects to HPC clusters for heavy bioinformatics workloads, and calls specialized foundation models for protein structure prediction and design. The agent itself decides which resource is needed and in what order. "The agent is the brain coordinating all these action types to perform end-to-end analysis while handling large data and managing machines," says Tianwei She, founding engineer at Phylo working on product, backend, agent orchestration, and the evaluation pipeline that determines quality.

LLM inference is the bulk of Phylo's cost base, and model choice dictates the product's cost structure. Biomni Lab is usage-priced with a limited free tier; token consumption in these agentic runs is especially high, and inference cost wasn't just a line item in the infrastructure bill — it set the usage quota, and the quota set how much science a biologist could do before hitting a paywall.

At the same time, researchers asked for control over the model running their work. Until recently, strong open-weight models were enough to offer a real alternative. "We started with more proprietary models. We wanted to bring Biomni Lab to as many scientists as possible at the same frontier quality while lowering costs. At the same time, our users really want the flexibility to choose models. There are so many good open-source models out there, and we wanted to give them that control," She says.

Self-hosting wasn't on the table — a small team tackling a hard scientific problem shouldn't burn engineering hours on a serving stack. Phylo evaluated several inference providers and ran a multi-week technical evaluation across multiple models on Fireworks. Three factors decided the partnership; the full details were not published in the truncated source.

Phylo runs its internal BiomniBench to measure how well each model handles these data-analysis tasks. Open frontier models scored high against this benchmark. The team's conclusion: open models are already sufficiently capable for a wide range of biomedical work, and efficient serving lets them put that capability in front of more scientists. This is a manufacturer claim based on an internal benchmark, not an independent third-party test, and should be read with that caveat.