Phylo is Reshaping Life Sciences with Open-Source AI: Costs Cut in Half, Research Accelerated

Overview

Phylo Labs recently announced that, in collaboration with Fireworks AI, it has successfully transformed the cost structure and efficiency of life sciences research by deploying cutting-edge open-source models on a server-side inference platform. As a derivative application of Stanford's open-source project Biomni, Biomni Lab aims to provide biologists with an integrated intelligent research environment. According to Phylo data, this technological shift has doubled the research output supported per dollar, while user scale has grown twofold month-over-month and inference costs have decreased by 60%.

"The agent is the brain driving all different operations, responsible for end-to-end analysis and handling massive data and machine scheduling." — Tianwei She, Founding Engineer at Phylo

This breakthrough marks a key step in the strategic transition from relying on proprietary models to open-weight models, enabling the scientific community to access the most advanced artificial intelligence capabilities at a lower threshold.

Challenges: High Inference Costs Constraining Research Accessibility

The core application scenarios of Biomni Lab involve long-cycle tasks such as literature reviews, hypothesis generation, experimental design, and data analysis. These tasks represent typical long-horizon agentic inference, where a single run may last hours or even days, requiring hundreds of tool calls and connections to high-performance computing clusters.

Phylo's primary challenge lay in the cost pressure caused by massive token consumption. In the early stages, Phylo mainly relied on proprietary large models. Since Biomni Lab operates on a pay-as-you-go model, inference costs directly determined users' usage quotas. With the rapid growth of its user base, a doubling of monthly usage meant that bills for proprietary models surged correspondingly, severely limiting more biologists' ability to access the platform. For startup teams, self-hosting models not only entailed high maintenance costs but also diverted R&D efforts; therefore, finding efficient external inference support became a necessity.

Solution: Model-Agnostic Architecture and Open-Source First

To address these challenges, Phylo designed a model-agnostic architecture and established the internal benchmark BiomniBench, led by founding engineer Tianwei She. This benchmark is specifically used to evaluate the quality, latency, and cost performance of various models when handling biological data analysis tasks.

Based on the evaluation results, Phylo configured default model options for each category of task and selected Fireworks as the primary host for open-source models. Fireworks' zero-data-retention serverless endpoints satisfied the security and privacy requirements for research data while also providing day-zero access to the latest open-source models—that is, making them available in production environments within 24 hours of their release. Currently, most of Phylo's traffic has migrated to open-source models, retaining only some proprietary models for specific high-difficulty tasks to ensure cutting-edge quality.

Results: Halved Costs and Accelerated Research

Optimizations in the technical architecture have brought significant business and research benefits. First, research output per unit cost has improved substantially, and higher utilization efficiency of user quotas has enabled more scientists to undertake long-term agentic inference tasks. Second, the introduction of open-source models has significantly reduced latency; Fireworks' fast path has nearly halved the time to first token, enhancing the fluency of interactions between biologists and agents.

Additionally, Phylo has retained users' control over model selection, meeting the needs of senior researchers who require interpretability or specific performance characteristics. Through this transformation, Phylo has not only verified the feasibility of commercial implementation of open-source AI in vertical domains but also provided a replicable template for the democratization of AI in the life sciences field. As the team expands, Phylo continues to deepen its collaboration with the open-source community, driving AI to become infrastructure for biologists' daily research.