Endless Frontier open-sources agentic model for biology research
The new Fx-Bio-1031 model allows researchers to run complex, multi-step bioinformatics analyses on local hardware while maintaining data privacy.
The short version
- Fx-Bio-1031 is an agentic model with 180 billion total parameters, built on Alibaba's Qwen3.8-Flash-Next.
- The model can autonomously write scripts, install software, and query databases like NCBI and Ensembl.
- It scored 70.59% Pass@5 on the Difficult subset of Anthropic's BioMysteryBench, beating GPT-6-Astra.
- The model supports a large context window of 524,288 tokens and is released under the Qwen Community License 1.0.
- It is designed for hospitals and drug developers to run analyses on local hardware without sharing data off-site.
Why it matters
This release provides a high-performing tool for scientists to automate tedious data cleaning and analysis tasks. Because it can run locally, it solves a major hurdle for medical and pharmaceutical companies that cannot share private patient data with cloud providers.
A new tool for biology
The Endless Frontier has officially released the full weights for Fx-Bio-1031, a specialized large language model for bioinformatics and computational biology. This model is categorized as an agentic model, meaning it is designed to plan and execute multi-step tasks independently rather than just generating text. It was built by post-training Alibaba's Qwen3.8-Flash-Next and is now available on Hugging Face.
The model is specifically engineered to handle the complexities of biological data analysis. It can perform actions such as writing necessary scripts, installing the required software, and querying established databases like Ensembl and NCBI. By checking its own results as it goes, the model acts more like a research assistant than a standard chatbot.
The release comes from a company incubated by Shanghai Jiao Tong University and DP Technology. These organizations aim to provide high-level tools for research groups and drug developers. The company previously released a general-purpose model called BigBang-V1 in August, but Fx-Bio-1031 focuses specifically on the life sciences.
Technical architecture and scale
Fx-Bio-1031 features a total of 180 billion language parameters. This total is composed of several distinct components, including a 125-billion-parameter mixture-of-experts backbone. A mixture-of-experts is a technique where only a small portion of the model's parameters are active at any given time, which helps manage computational efficiency.
In addition to the backbone, the model includes a 51-billion-parameter n-gram embedding and a 4-billion-parameter multi-token prediction module. These components work together to allow the model to process information more effectively. The released configuration supports a context window of 524,288 tokens, allowing it to handle very long sequences of data in a single session.
The model weights are approximately 360GB in BF16 format. These weights are published under the Qwen Community License 1.0. While the weights are open, the company directs commercial users to specific terms within that license to understand how they can use the model for business purposes.
Performance on biology benchmarks
The developers evaluated Fx-Bio-1031 using an agentic evaluation framework across several benchmarks. On the Difficult subset of Anthropic's BioMysteryBench, which consists of problems that domain experts could not solve alone, the model achieved a 70.59% Pass@5 score. This outperformed GPT-6-Astra, which scored 47.06%, and Opus-5, which scored 52.94%.
On the expert-solvable subset of the same benchmark, the model scored 93.15%. This was slightly lower than the 94.52% achieved by Opus-5. In another test called BiomniBench, Fx-Bio-1031 reached an 89.50% Pass@3, which was ahead of Opus-5's 87.42% score. These results suggest the model is particularly strong at tackling complex, non-obvious biological problems.
The company noted that no public benchmarks were used during the training or data synthesis phases. This means the model's performance on these tests is a reflection of its learned capabilities rather than memorized answers. The developers also tested it on BixBench3, where it scored 45.76, placing it above Opus-5 but below GPT-6-Astra.
How the model was trained
The training process utilized a pipeline the company calls RSI-OS. This is a recursive self-improvement pipeline that covers evaluation, training data, and the tool environments the model interacts with. By creating a loop where the model can be evaluated and improved against its own tools, the developers aimed to create a more capable agent.
To refine the model, the team built an internal test set of 114 complex tasks. They collaborated with DP Technology's BioMaster team to develop these tasks. This internal testing helped ensure the model could handle the specific nuances of bioinformatics, such as navigating complex database queries and interpreting biological mutations.
One specific example of the model's capability involves a breast cancer cohort task. The model took 1,918 raw records and cleaned them down to 762 unique patients. It then identified that certain mutations and pathway alterations were mutually exclusive and ran 14 sensitivity analyses with negative controls to map where those conclusions held true.
Impact on private research
A primary goal for Fx-Bio-1031 is to serve hospitals, research groups, and drug developers. These entities often handle highly sensitive data that cannot be uploaded to public cloud services due to privacy regulations. By providing the full weights, Endless Frontier allows these organizations to run the model on their own local hardware.
This local deployment is a significant advantage for medical research. It ensures that patient data remains on-site while still giving researchers access to advanced AI analysis. The model's ability to write its own scripts and query databases means researchers can automate the most time-consuming parts of their data pipelines.
Moving forward, the availability of such a model could accelerate drug discovery and genomic research. It provides a way to bridge the gap between raw biological data and actionable insights. Researchers can now use an agentic system to perform multi-step analyses that previously required manual coding and significant human oversight.
What to watch
- How the model performs on real-world, non-benchmark biological datasets.
- Whether other companies release similar agentic models for specialized scientific fields.
Try it
Check out the Fx-Bio-1031 model card on Hugging Face to see the specific tasks it was tested on.
Sources and links
- Reportingpandaily.com: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by pandaily.com · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
