Endless Frontier open-sources Fx-Bio-1031: a bioinformatics agent that clears 70.59% of expert-unsolved BioMysteryBench tasks
Synopsis
The Endless Frontier lab has post-trained Fx-Bio-1031, an agentic model for bioinformatics and computational biology, on Qwen3.8-Flash-Next and released the full weights. Rather than static question answering, it plans and executes multi-step tasks such as multi-omics analysis and pipeline coding and debugging inside sandboxed environments. On Anthropic's BioMysteryBench, its Pass@5 on the Difficult subset, problems domain experts failed to solve, is 70.59 against 52.94 for Opus-5 and 47.06 for GPT-6-Astra; on the Solvable subset it reaches 93.15, below Opus-5's 94.52, and its BixBench3 score of 45.76 trails GPT-6-Astra's 48.70.
Interpretation
Bioinformatics agents now have a dedicated model with open weights that can autonomously carry out multi-step tasks in an interactive execution environment, including multi-omics data analysis, pipeline coding and debugging, gene-function investigation and integration of evidence from multiple sources. Earlier models stayed largely at static question answering and did not take on planning and execution across steps; Fx-Bio-1031 shifts the mode of work to calling real tools and data inside sandboxed containers to complete an entire analysis chain. On Anthropic's BioMysteryBench, Fx-Bio-1031 scores 70.59 Pass@5 on the Difficult subset that domain experts could not solve independently, against 52.94 for Opus-5 and 47.06 for GPT-6-Astra; its Overall Pass@5 is 88.89, above Opus-5's 86.67 and GPT-6-Astra's 83.33.
The model is post-trained on Qwen3.8-Flash-Next, with training data drawn from execution-verified agentic trajectories covering tool-augmented biological data analysis and multi-step scientific workflows. The training data is not a fixed corpus but is iterated through trajectory curation, execution verification, evaluation and training feedback, so its scale and quality keep changing; Fx-Bio-1031 is one checkpoint in that process. Evaluation spans three bioinformatics agent benchmarks, BioMysteryBench, BiomniBench and BixBench3, with all evaluated models run under the same internal agentic evaluation framework.
On the process-level BiomniBench, Fx-Bio-1031 reaches 89.50 Pass@3 and 82.46 Avg@3, both above Opus-5's 87.42 and 81.59 and GPT-6-Astra's 82.00 and 76.88. That benchmark grades the whole analytical trajectory on real experimental data, from data cleaning and method selection through statistical testing and biological interpretation, rather than only the final answer. In the same results table, Fx-Bio-1031's BixBench3 Score is 45.76, below GPT-6-Astra's 48.70 and above Opus-5's 40.60.
Perspective
The demonstrated scope is evaluation on three bioinformatics agent benchmarks, all obtained under the same internal agentic evaluation framework; the release includes full weights, tokenizer, processor configuration and generation configuration under the Qwen Community License 1.0.
The releaser describes Fx-Bio-1031 as one checkpoint in an ongoing evolution rather than a fixed endpoint, with training data scale and quality still changing; leadership is not uniform across benchmark subsets, and the BixBench3 score is below GPT-6-Astra's.
