Skip to content
AI.info

The Pulse

Endless Frontier releases weights for 180B bioinformatics model

Fx-Bio-1031 is designed to carry out multi-step bioinformatics analyses using tools and biological data. The lab reports benchmark results from its own evaluation framework, with mixed comparisons against other models.

Endless Frontier releases weights for 180B bioinformatics model

The Endless Frontier has released the weights of Fx-Bio-1031, a 180-billion-parameter model designed to carry out multi-step bioinformatics analyses. The lab's model card says it was post-trained from Qwen3.8-Flash-Next for tasks including coding and debugging pipelines, investigating gene functions and combining evidence from multiple sources.

AI Atlas dates the release to Nov. 13. The model is intended to plan and execute tasks in interactive environments rather than only answer questions.

From questions to workflows

The model was post-trained on curated examples of tool-assisted biological data analysis and multi-step scientific work. According to AI Atlas, its intended activities include writing scripts, installing software, querying databases such as NCBI and Ensembl, and checking results.

The release includes the complete weights, tokenizer, processor configuration and generation configuration. The model card also describes a multimodal causal language model with a vision encoder.

The lab describes its training process as recursive self-improvement, using iterative curation of task trajectories, execution verification, evaluation and training feedback. It presents this model as a checkpoint in an ongoing development process.

Model size and access

The 180 billion language parameters comprise a 125-billion-parameter mixture-of-experts backbone, with 6 billion activated, along with a 51-billion-parameter n-gram embedding and a 4-billion-parameter multi-token prediction module. The configuration has 512 routed experts, with 10 routed experts and one shared expert active per token.

The released configuration supports a context length of 524,288 tokens. The weights are in BF16 safetensors format, split across 73 shards, with approximately 360 GB of tensor data.

The model is distributed under the Qwen Community License 1.0. The model card provides an example of using an OpenAI-compatible API with a locally served model, as well as references to serving engines.

Benchmark results

The Endless Frontier reports results from its internal agentic evaluation framework across the models it tested. On the Difficult subset of BioMysteryBench, Fx-Bio-1031 scored 70.59% Pass@5, compared with 52.94% for Opus-5 and 47.06% for GPT-6-Astra.

The model card describes the Difficult subset as verifiable scientific problems that domain experts had not solved independently. On the Solvable subset, the lab reports 93.15% Pass@5 for Fx-Bio-1031, below Opus-5's 94.52%.

On BiomniBench, the reported Pass@3 result was 89.50% for Fx-Bio-1031, against 87.42% for Opus-5. The card describes that benchmark as grading an agent's analytical trajectory, including data cleaning, method selection, statistical testing and biological interpretation.

On BixBench3, Fx-Bio-1031 scored 45.76 in the lab's evaluation, compared with 48.70 for GPT-6-Astra and 40.60 for Opus-5. The model card says the benchmark tests agents on biological data and multi-step workflows in interactive notebook environments.

Sources

Related stories

Explore

More articles