Phylo Brings Frontier Ai To More Scientists With Open Models On Fireworks

Phylo brings frontier AI to more scientists with open models on Fireworks

Join us for our inaugural conference, Forge 2026

Blog

Phylo Brings Frontier AI To More Scientists With Open Models On Fireworks Phylo brings frontier AI to more scientists with open models on Fireworks

PUBLISHED 9/17/2026

Table of Contents At-a-glance Agents that run for hours to days From proprietary models to open weights Fast models, fast answers, no friction

Day-zero model access

Direct engineering access

Developer experience that held up in production Frontier quality, more science per dollar

More work per quota, better retention

Latency improved alongside cost

Open models already excel at biomedical work From inference to training Advice for founders: Spend your time on the problem only you can solve

Table of Contents

Start using Fireworks

Sign up

At-a-glance

Industry Life sciences research and AI for science Use case Agentic integrated biology environment supporting the scientific community in research planning, execution, and analysis Workload Long-horizon agentic inference. Hundreds of tool calls per run, executing for hours or days Models Frontier open-weight models, with proprietary models retained for a subset of tasks Deployment Fireworks fast path serverless endpoints with zero data retention

Challenge • Token-hungry workloads. Long-horizon agentic runs execute for hours or days, scaled across multiple machines. • Cost capped the science. Inference spend set the usage quota, which limited how much work a biologist could do before hitting a paywall. • Rapid adoption, rising load. Doubling monthly usage meant proprietary model costs would rise in step, limiting how far the product could reach.

Solution • Model-agnostic by design. Defaults set for each task type by Phylo's own BiomniBench evaluations for quality, latency, and cost. • Open weights on Fireworks. Serverless inference with day-zero access to state-of-the-art open models. • One day to production. Automated evals, live traffic test, add to production.

Results • 2x month-on-month user growth at 60% lower cost. More scientific work per quota drives higher retention. • Day zero access. New state-of-the-art open models in production within 24 hours, giving biologists the fastest access to frontier-level intelligence. • Majority of traffic now on open models , with time to first token roughly halved on Fireworks serverless fast path.

Phylo is an applied research lab building Biomni Lab, the first integrated biology environment. It grew out of Biomni, an open-source project started at Stanford in 2024 and now used by more than 50,000 scientists in labs around the world. Biomni Lab gives biologists agents that plan, execute, and document research across hundreds of integrated databases and tools, with scientists reporting up to 40x faster hypothesis-to-analysis cycles. Biomni Lab was built by a small founding team ahead of public launch. Phylo now numbers dozens of team members and is growing fast . Tianwei She, a founding engineer, has been there since day one and works across the product surface, backend, agent harness, and the evaluation pipeline that governs agent quality. LLM spend is the majority of Phylo's cost base, so model selection and pricing strategy are handled together as one decision: choosing a model sets the product's cost structure. Biomni Lab is model-agnostic by design. Phylo evaluates every proprietary and open-weight frontier model, then configures a default for each task type against its own benchmarks for quality, latency, and cost. Users also have access to a custom model selector, where they can select the specific model they prefer. Fireworks serves the open-weight side of that selection. Agents that run for hours to days

Biologists use Biomni Lab across literature review, ideation, hypothesis generation, experimental design, and data analysis. These are long-horizon agentic runs with a single task executing for hours or days against datasets ranging from tens to hundreds of gigabytes. The agent spins up multiple machines, connects to HPC clusters for high-compute bioinformatics work, and calls specialized bio foundation models for protein design and structure prediction. The agent decides which of those resources a given task needs, and in what order.

"The agent is the brain that facilitates all these different types of operations to conduct analysis end to end, handling the large data and machine management." Tianwei She, Founding Engineer, Phylo

Phylo runs its internal BiomniBench benchmarks to measure how well each model handles these data analysis tasks. Frontier open-weight models scored strongly against it. Open models are already sufficiently capable for a wide range of biomedical work, and serving them efficiently means Phylo can put that capability in front of more scientists.

From proprietary models to open weights

Phylo started on proprietary models. Two forces changed the architecture. The first was economics . Biomni Lab is priced on usage, with a free tier carrying a usage limit. Agentic biology consumes tokens at a rate few workloads match, and the user base is growing fast. Inference cost was not a line item on an infrastructure bill. It set the usage quota, and the quota set how much science a biologist could do before hitting a paywall. The second was user demand . Biologists wanted control over which model ran their work. By then, enough strong open-weight models existed to give them a genuine choice. "We started using more proprietary models in the beginning. We wanted to bring Biomni Lab to as many scientists as we could at the same frontier quality while driving cost efficiencies. At the same time, our users really want the flexibility of choosing models. There are so many good open source models out there, so we wanted to give users that control." Tianwei Self-hosting models was never on the table. With a small team building against a hard scientific problem, running a serving stack was not where the engineering hours belonged. Phylo assessed multiple inference providers, then ran a multi-week technical evaluation across several models on Fireworks. With Biomni Lab, biologists can execute complex tasks in minutes that normally take weeks. Design proteins. Analyze multi-omics data. Read and summarize thousands of papers. Fast models, fast answers, no friction

Three factors cemented the partnership with Fireworks: how fast Phylo could get new models...