← All model news

Molence Bio publishes Mol-1 code and weights: self-reported 75% antibody-antigen structure prediction success, single-GPU agent assembly for very large complexes

According to an October 31, 2026 report by Xinzhiyuan, Chinese AI drug discovery company Molence Bio released the biomolecular foundation model Mol-1 and published its code and weights on GitHub. The company reports a 75% success rate for antibody-antigen structure prediction; very large complexes are assembled by a long-running agent on a single H20; the nanobody design data come from an unreleased earlier internal version.

Published Updated

What changed

Mol-1 is Molence Bio's first biomolecular foundation model. The company says a single model covers antibody-antigen, protein-peptide, protein-small molecule and molecular glue complex structure prediction, and it offers two versions, Mol-1 and Mol-1-constraint with epitope constraints. According to figures in the report: On the FoldBench-AB antibody-antigen test set, Mol-1 has a 75.0% success rate, against 75.6% for IsoDDE in Isomorphic Labs' technical report; with epitope constraints it rises to 88.3%. The company also enumerated 201,376 five-seed combinations from 32 random seeds; 62.22% reached at least 75% success and 17.85% exceeded 76%. For peptide-protein complexes, the company built MolBench-Peptide-26, covering 172 interface clusters from 150 PDB entries published in the first three quarters of 2026; at the strictest DockQ>0.80 threshold the success rate is 54.1%, which the company says compares with 52.3% for the best-performing unconstrained control and 60.2% for Mol-1-constraint. On the highly homologous molecular glue subset the success rate is 74.6%, with the reported second place at 40.7%. For very large complexes, a long-running agent driven by Mol-1 evaluated 10 large complexes on a single NVIDIA H20; 9 beat corresponding results previously reported by Anthropic, and 8 had TM-score above 0.9; the largest system had 60,000 residues and 120 chains, with TM-score 0.997. The comparison is Anthropic's public work using an eight-B300-GPU node and a maximum accurate prediction of about 10,000 amino acids. De novo nanobody design comes from an internal platform, and the report states it is based on an early version of Mol-1: binding candidates were obtained for all 17 targets, with the highest per-target hit rate at 63%; for BCMA, 12 of 30 VHH candidates bound, and the best candidate measured KD 43 pM by SPR; for 4-1BB agonist VHHs, after one round of experimental feedback, 51 of 60 second-round designs had functional activity, raising the hit rate from about 7% to 85%; an antagonist VHH against CXCR6 had EC50 39 nM and inhibited ligand signaling; for two antigens with only 54% sequence identity, 5 of 30 first-round designs bound both. The report says the code and foundation model weights are public on GitHub, and online prediction services for Mol-1 and Mol-1-constraint are also offered; local deployment can support more complex prediction, design and agent workflows.

What it means for developers

This kind of model is integrated very differently from a chat model: inputs are sequences and structures, outputs are coordinates, and acceptance is judged by structural metrics such as DockQ and TM-score, not text scores or human readability. So whether you can put it into production depends on whether you have a local holdout set with experimental structures. The most direct value of public weights is that target sequences can stay on the internal network and batch complex runs do not have to queue for a web service. But public weights are not the same as a license grant; the report does not state the license, so commercial use and redistribution require checking the repository. Two points are genuinely informative for developers. First, constrained mode brings a consistent gain (antibody-antigen from 75.0% to 88.3%, peptide complexes at DockQ>0.80 from 54.1% to 60.2%), which shows whether epitope or interface residue annotations are available directly changes result quality. That affects your input-preparation workflow more than model selection. Second, very large complexes rely on a long-running agent that repeatedly predicts and assembles, and a single H20 achieved the reported comparison; this suggests the scheduling and orchestration layer may be more cost-effective than switching to a larger model or adding GPUs. These numbers need to be discounted: most comparisons in the report use company-built datasets or cite figures from third-party technical reports, not same-environment reruns; the 75% figure is sensitive to random seeds, with 17.85% of five-seed combinations exceeding 76% and, conversely, about 38% below 75%; the design section comes from an early internal version and internal wet-lab experiments, and there is no evidence that the public weights can reproduce those hit rates.

Integration and evaluation

The following are conditional recommendations, not facts from the report. First, decide whether to self-host or start with the online service. The reported change is that code and base weights are public while the web service remains. If target sequences cannot leave the internal network, or you need batch complex runs and want to chain your own agent, self-hosting is the only option; if you only occasionally check one or two structures, the web service is less effort. For the decision, use your own holdout set: pick 20 to 50 antibody-antigen or peptide-protein complexes with experimental structures, compute the success rate by DockQ or the company's definition and compare with 75%, and record GPU memory use and runtime per complex. If reproduction is clearly lower, suspect version, input preprocessing and evaluation-script differences first rather than rejecting the model outright. Second, treat constrained mode as a switch with preconditions, not a default. The reported gains are FoldBench-AB 75.0% to 88.3% and peptide complexes at DockQ>0.80 from 54.1% to 60.2%, and the precondition is that you have epitope or interface residue annotations; without annotations this path is unavailable, and forcing it amounts to introducing a wrong prior. A practical acceptance test: compare constrained and unconstrained DockQ>0.80 pass rates on your own peptide-protein set; if the gap is much smaller than reported, do not make the constraint a hard gate in the pipeline. Third, for very large complexes, try orchestration before adding GPUs. The reported comparison is eight B300 GPUs versus a single H20, and the difference comes from the agent repeatedly predicting and assembling, not from model scale; the trigger condition is systems reaching tens of thousands of residues and dozens of chains. Acceptance metrics can be: on the 5 to 10 large complexes you care about, count how many reach TM-score 0.9 and total wall-clock time, and compare with one-shot direct prediction; if direct prediction is sufficient, the agent only adds cost. The precondition is to confirm whether the repository includes this orchestration code; the report only says code and weights are public and does not itemize the agent part. Fourth, do not treat nanobody design as a delivered capability. The hit-rate and affinity data come from an internal platform and an early model version; what is public is the structure prediction model. If your route depends on de novo VHH design, you can first use Mol-1 as a structure-validation component: on a candidate set with a known affinity gradient, see whether predicted structures can rank them in the correct order, then decide whether to connect it into the design loop.

Limits and open questions

Mol-1 is not currently in KimiSeek's model catalog; the online service and weight download mentioned in the report refer to Molence Bio's offerings, and this site does not provide access to the model. The report does not give parameter count, minimum GPU memory, inference dependencies or license, so the questions "can it run on my own GPU" and "can it be used commercially" cannot currently be answered; only the repository and official pages can settle them. All accuracy figures are company-reported: some comparisons are cited from third-party technical reports rather than rerun in the same environment; MolBench-Peptide-26 was built by the company; the very-large-complex comparison is against results Anthropic previously published, not a parallel evaluation on the same set of systems. The nanobody section is wet-lab results from the company's internal laboratory and is explicitly based on an early version of Mol-1, not the same object as the public weights, so it cannot be used to infer the design capability of the public version. The report date is October 31, 2026; repository contents, license terms and service availability may all change afterward.

Three minutes to your first working call

Create an account and an API key. Calls use your account balance. 30% off your first top-up. Top up ¥500, pay ¥350, receive ¥500 in credit. Top up ¥1000, pay ¥850, receive ¥1000 in credit You can still start with a ¥10 trial top-up.

Sources and editorial process

追平DeepMind、反超Anthropic!国产开源模型闯入AI制药第一梯队

AI assisted the analysis and translation of the cited sources, followed by automated quality checks. Recommendations are not test results. Consult the current catalog for availability and pricing.