Endless Frontier open-sources Fx-Bio-1031, a 125B MoE bioinformatics model with 512K context

Endless Frontier (无界前沿) released the full weights of Fx-Bio-1031 on November 13, 2026, a bioinformatics model with 125B total parameters and 6B activated per token. According to the company, it scores 70.59% Pass@5 on the Difficult subset of Anthropic's BioMysteryBench, ahead of GPT-6-Astra (47.06%) and Claude Opus-5 (52.94%) under the same framework. The announcement was reported by Chinese outlet QbitAI (量子位) in a Chinese-language article, and the benchmark figures below are the company's own.

Specifications

• Architecture: Mixture-of-Experts, 125B total parameters, 6B active per token
• Base: post-trained from Qwen3.8-Flash-Next using what the company calls RSI (Recursive Self-Improving) post-training
• Context window: 512K tokens, native, with a thinking mode
• License: Qwen Community License 1.0; commercial use requires following the license's commercial terms
• Weights: full weights, tokenizer, processor config and generation config on Hugging Face (endless-frontier/Fx-Bio-1031)
• Pricing: no API pricing stated

Benchmark claims

On BioMysteryBench Difficult, Fx-Bio-1031 reached 70.59% Pass@5. The company says this is 23.53 points above GPT-6-Astra and 17.65 points above Claude Opus-5. Pass@k is the share of tasks solved at least once in k attempts. On BiomniBench, it reports 89.50% Pass@3. The source credits the comparison chart to Hugging Face's official evaluation.

BioMysteryBench tasks place the model in real biological data and a Linux tool environment. They cover whole-genome sequencing, scRNA-seq, ATAC-seq, metagenomics and proteomics.

The company says public benchmarks including BioMysteryBench and BixBench3 were used only as final blind tests, with no role in training, fine-tuning or data synthesis.

Example tasks from the company

• BiomniBench da-18-7: from 1,918 breast cancer sample records, the model deduplicated to 762 patients. It reported an RAS/MAPK pathway variant rate of 7.1% in ESR1-mutant patients against 19.5% in wild-type (OR=0.317, p=0.0011). It then ran 14 sensitivity analyses with PIK3CA and CDH1 as negative controls. The exclusivity effect held under most strict definitions but lost significance when variants were defined too broadly or only amplifications were counted.
• BioMysteryBench hb047: nine unlabeled hiPSC ATAC-seq samples with 110,559 peaks. The model clustered them into three groups using correlations across 83,354 loci, then used public HDAC inhibitor response data to identify the treated and control groups.
• Mutational signature task: 2,700 tumor samples and 96 mutation channels. The model fit COSMIC SBS signatures with NNLS, cross-checked with official tools, and ran 200 bootstrap resamples over 76 steps.
• Bacterial flagellar H antigen typing (BioMaster platform): the model found a malformed final record in an R2 sequencing file, removed it, and ran SeqSero2. It finished in 258 steps over 34.5 minutes.

Training approach

Endless Frontier describes a three-layer RSI loop it calls RSI-OS: evaluation, data and environment. For evaluation, it worked with the BioMaster team from DP Technology (深势科技) on 114 internal tasks, run three times for 342 task attempts and more than 20,000 model requests. The data layer has AI help generate, solve and review training tasks. The environment layer has the model help write and repair analysis scripts and the agent harness. The company's BigBang-V1, released in August, applied the data-layer approach to a 35B model. Training used DP Technology's Trisol machine learning platform.

Endless Frontier was initiated by a Chinese Academy of Sciences academician and incubated by Shanghai Jiao Tong University, the Shanghai Academy of Algorithm Innovation (上海算法创新研究院) and DP Technology. Core members Chen Siheng and Zhang Linfeng previously led the Innovator and SciMaster projects.

What this means

Fx-Bio-1031 makes a specialized bioinformatics model available as open weights. With 6B active parameters, the company says it can be deployed on local hardware, keeping clinical and omics data on premises. The reported results come from the company's agent evaluation framework, and the comparison with GPT-6-Astra and Claude Opus-5 rests on the company's runs. The Qwen Community License 1.0 governs commercial use, so teams planning commercial deployment need to review its terms before building on the model.