The Endless Frontier open-sources Fx-Bio-1031 bioinformatics agent
The Endless Frontier has released open weights for a 180-billion-parameter bioinformatics agent, but its benchmark claims still need independent scrutiny.
Fx-Bio-1031 is an open-weight agentic model aimed at bioinformatics and computational biology. The Endless Frontier says it can plan and execute multi-step work including multi-omics analysis, bioinformatics pipeline coding and debugging, gene-function investigation and evidence integration in interactive environments. The complete weights, tokenizer and configuration files are available through the model card. 1
What it is
The release is built on Qwen3.8-Flash-Next and is listed as a 180-billion-parameter model. That headline number includes a 125-billion-parameter mixture-of-experts backbone, with roughly 6 billion parameters active per token, alongside additional n-gram embedding and multi-token-prediction components. It supports a 524,288-token context window and the published BF16 files occupy roughly 360GB across 73 shards. 1
That scale makes this a serious infrastructure project rather than a model most developers will casually run on a workstation. The model card says it is released under Qwen Community License 1.0, so anyone considering deployment will need to read the licence and work out whether their intended use fits its terms. 1
The reported results
The Endless Frontier reports a 70.59 Pass@5 result on the Difficult subset of BioMysteryBench, 93.15 on its Solvable subset and 88.89 overall. It also reports 89.50 Pass@3 on BiomniBench and a 45.76 score on BixBench3. In the comparison table, Fx-Bio-1031 leads the listed Opus-5 and GPT-6-Astra figures on several measures, although GPT-6-Astra remains ahead on BixBench3. 1
Those numbers are developer-reported results from the lab's internal evaluation framework. They are not independent validation, clinical evidence or permission to put unreviewed model output anywhere near patient data. The model card itself does not establish medical approval or safe operation in a clinical setting. 1
Our opinion
Fx-Bio-1031 is notable because the release pairs a specialised scientific-agent pitch with genuinely open weights, rather than offering only a hosted demo. The catch is equally clear: a 360GB download and the need for careful scientific verification put it far beyond casual experimentation. The benchmark chart is a useful signal that the team has targeted difficult workflows, but it is still a claim to investigate, not a substitute for reproducible external testing. For researchers with the hardware and expertise to audit it, this is an interesting release; for everyone else, the model card is more useful as a map of where open scientific agents are heading than as a turnkey tool.