Open-Weight Model Trained on Lab Records Tops GPT-6

Researchers gain an MIT-licensed research model whose reported results, still unverified independently, point to AI trained on the working records of experiments.

Updated on Nov. 3, 2026 in Chemistry

Open-Weight Model Trained on Lab Records Tops GPT-6

Deep Principle has released Hydrogen-1.0, a research model built on its GLM-5.3 base, with model weights available under the MIT license. The company reports a 38.3% pass rate on OpenAI's FrontierScience Research evaluation, ahead of 33.3% for GPT-6 Astra; the work is research-stage, not a shipped product.

Why it matters

The release points toward AI systems trained on the working records of laboratory science, such as experiment design, errors and mid-course changes, rather than only on published papers. Making the weights open lets outside groups test the claims directly.

The figures are company-reported comparisons on a single research evaluation: Hydrogen-1.0 posted a 38.3% pass rate against 33.3% for GPT-6 Astra on OpenAI's FrontierScience Research test, and its best-so-far HOMO-LUMO gap of 4.318 eV compares with 3.689 eV for GPT-6-Astra and 3.353 eV for Kimi-K3. Independent replication and the complete training data have not been made available.

The players

Deep Principle

An AI research developer that built Hydrogen-1.0 on the GLM-5.3 base model and released its weights under the MIT license.

OpenAI

The AI lab that runs the FrontierScience Research evaluation on which Hydrogen-1.0 and GPT-6 Astra were compared.

The details

Deep Principle trained Hydrogen-1.0 on laboratory records describing how experiments were designed, where they went wrong and what researchers changed along the way, material that scientific papers usually omit. In one test, the model detected and repaired a topology error in a polyethylene melt simulation and restarted the run. In a virtual lab rehearsal, it wrote a JSON protocol for a 12-vial Suzuki coupling study and corrected its own sequence after catching a step that would have used a used pipette tip.

Timeline

  1. Article on Hydrogen-1.0 results and release published November 3, 2026.

  2. Deep Principle ran five rounds of new-material screening, each informed by the previous round's evaluation.

The Tech Race

Hydrogen-1.0's 38.3% pass rate on OpenAI's FrontierScience Research evaluation edges out GPT-6 Astra's 33.3%, placing an open-weight research model against a leading closed frontier system on the same test. The comparison is drawn from Deep Principle's own reporting and has not been independently verified.

Researchers can download the Hydrogen-1.0 weights under the MIT license and run them on their own hardware, but the full training datasets and recipes were not released. No availability date for a hosted product or a commercial release has been announced.

The takeaway

Watch for independent replications of the FrontierScience Research pass rates and whether the screening results hold outside Deep Principle's own evaluations. Any released training data or recipes would show how far the laboratory-record approach can be reproduced.

Further reading

For more on the chemistry side of this work, see the Chemistry section.

Source note: This article includes information reported by TokenPost.