Fx-Bio-1031

Agentic Bioinformatics & Computational Biology through RSI-Driven Continuous Improvement

_{Qwen3.8-Flash-Next  ·  524K Context  ·  Open Weights}

---

Overview

Fx-Bio-1031 is an agentic large language model for bioinformatics and computational biology, post-trained on Qwen3.8-Flash-Next by The Endless Frontier lab. Moving beyond static question answering, it is designed to autonomously plan and execute complex, multi-step biological tasks—including multi-omics data analysis, bioinformatics pipeline coding and debugging, gene-function investigation, and multi-source evidence integration within interactive execution environments.

The model is post-trained on curated agentic trajectories covering tool-augmented biological data analysis and multi-step scientific workflows. This release includes the complete model weights, tokenizer, processor configuration, and generation configuration.

Continuous improvement through Recursive Self-Improvement (RSI)
Fx-Bio is developed through an RSI-driven continuous improvement process. Our training data is never static: its scale and quality are constantly evolving through iterative trajectory curation, execution verification, evaluation, and training feedback. Each model iteration builds on these improvements, enabling sustained gains in agentic biological task execution over time. Fx-Bio-1031 is one checkpoint in this ongoing evolution, not a fixed endpoint.

Main Results

We evaluate Fx-Bio-1031 on three premier, industry-defining agentic benchmarks for bioinformatics and computational biology:

• BioMysteryBench (Dataset): Anthropic's flagship and sole official bioinformatics benchmark. Operating in sandboxed containers with real, anonymized multi-omics datasets and bioinformatics tools, it uniquely evaluates models not only on expert-verified solvable tasks (Solvable), but also on verifiable scientific problems that domain experts failed to solve independently (Difficult), testing the frontier of autonomous scientific discovery[cite: 1].
• BiomniBench (Paper · Dataset): The world's first process-level evaluation benchmark for biological research agents, built upon the Science-published Biomni ecosystem by Stanford and Phylo alongside over 100 senior domain experts. Its tasks are curated directly from real-world experimental datasets published in Nature*, Science, and *Cell, rigorously grading the agent's entire analytical trajectory—from data cleaning and methodology selection to statistical testing and biological interpretation.
• BixBench3 (Paper · GitHub · Dataset): FutureHouse's gold-standard benchmark for LLM-based agents in computational biology. It tests an agent's end-to-end ability to navigate real-world biological data capsules, author and execute multi-step Python/R/Bash workflows inside interactive notebook environments, and verify nuanced scientific hypotheses.

_{Comparison with GPT-6-Astra and Opus-5 on BioMysteryBench, BiomniBench, and BixBench3.}

Results are obtained using our internal agentic evaluation framework across all evaluated models. Higher is better; the best result in each row is shown in bold[cite: 1].

Benchmark	Metric	GPT-6-Astra	Opus-5	Fx-Bio-1031
BioMysteryBench	Solvable, Pass@5	91.78	94.52	93.15
BioMysteryBench	Difficult, Pass@5	47.06	52.94	70.59
BioMysteryBench	Overall, Pass@5	83.33	86.67	88.89
BiomniBench	Avg@3	76.88	81.59	82.46
BiomniBench	Pass@3	82.00	87.42	89.50
BixBench3	Score	48.70	40.60	45.76

Model Details

Base model	Qwen3.8-Flash-Next
Architecture	Qwen4ExpForConditionalGeneration
Model type	Multimodal causal language model with a vision encoder
Language parameters	180B total (125B MoE backbone with 6B activated, plus 51B n-gram embedding and 4B MTP)
Hidden layers	48
Experts	512 routed; 10 routed experts and 1 shared expert active per token
Context length	524,288 tokens in the released configuration
Weights	BF16 safetensors · 73 shards · approximately 360 GB of tensor data
Checkpoint	20261009-r2/iter_0001384
License	Qwen Community License 1.0

Quickstart

Use a recent inference framework with explicit support for Qwen3.8-Flash-Next / Qwen4Exp. For production workloads and high-throughput agentic execution, we recommend a dedicated serving engine:

• SGLang — Qwen3.8-Flash-Next cookbook
• vLLM — Qwen3.8-Flash-Next recipe

OpenAI-compatible API example

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="EMPTY",
)

response = client.chat.completions.create(
    model="endless-frontier/Fx-Bio-1031",
    messages=[
        {
            "role": "user",
            "content": (
                "Write a Python workflow to load an RNA-seq raw count matrix (counts.csv) "
                "and sample metadata (metadata.tsv), filter out low-expression genes, "
                "run differential expression analysis, and identify the top 5 candidate "
                "genes associated with the perturbation phenotype."
            ),
        }
    ],
    temperature=1.0,
    top_p=0.95,
    extra_body={"top_k": 20},
    max_tokens=4096,
)

print(response.choices[0].message.content)