Breaking

Deep Principle Releases Open-Weight Science Model Hydrogen-1.0, Claiming Lead Over GPT-6-Astra on Two Benchmarks

November 3, 2026 at 02:46 EST

Foundation Models
Open Source
Research & Papers

Deep Principle has released Hydrogen-1.0, an open-weight science model that the company says was post-trained on GLM-5.3 using laboratory records and scientific discovery trajectories.

Deep Principle · Hydrogen-1.0 · Open-weight science model

A science model trained on lab records and discovery trajectories claims a lead on research-level benchmarks

Deep Principle reports that Hydrogen-1.0, post-trained on GLM-5.3, outscores GPT-6-Astra on FrontierScience-Research. These are company-reported figures without independent evaluation or a reproducible technical paper.

38.3%

Hydrogen-1.0 on FrontierScience-Research

33.3%

GPT-6-Astra on FrontierScience-Research

93.0%

Hydrogen-1.0 on GPQA-Diamond, a graduate-level science benchmark

FrontierScience-Research: the 5-point gap, drawn to scale

Each block = 5 percentage points. Both models measured on the same benchmark, as reported by Deep Principle.

38.3%

Hydrogen-1.0

33.3%

GPT-6-Astra

Reported gap: 5.0 percentage points. Small margins like this are sensitive to prompting, evaluation settings and possible test-set contamination.

How Hydrogen-1.0 was trained

Company-described pipeline, not independently verified.

Laboratory records

→

Step-by-step discovery trajectories

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.