Deep Principle Turns GLM-5.3 Into Hydrogen-1.0, an MIT Scientific MoE Aimed at Lab Work and Materials Campaigns

Deep Principle has published Hydrogen-1.0, a scientific research model built by post-training GLM-5.3 on chemistry, materials science, and life-science data, together with laboratory records and discovery trajectories from its own labs. The company framed the November 2, 2026 release as a step toward open-weight systems that can help scientists formulate questions, design experiments, and revise plans from evidence—capabilities it says are still largely locked inside proprietary stacks.

The Hugging Face card lists a 753-billion-parameter mixture-of-experts checkpoint based on zai-org/GLM-5.3, a one-million-token context window, and a 128,000-token maximum output length. BF16 weights ship under the MIT License. Separate NVFP4 and MXFP4 quantized builds target NVIDIA Blackwell and AMD Instinct MI350-class hardware. Full training datasets and recipes are not released with the weights.

Why Deep Principle built another scientific checkpoint

The launch blog argues that recent high-profile scientific wins—OpenAI’s reported Navier–Stokes work with Lean formalization, Anthropic’s enzyme investigation—show models can help researchers explore hard problems, but the models behind those examples are proprietary. Hydrogen-1.0 is Deep Principle’s attempt to put comparable discovery-oriented behavior into inspectable open weights that labs can adapt.

Post-training draws on four source categories the company names explicitly: domain knowledge across chemistry, materials, and life sciences; task-specific feedback from its scientists; full-cycle electronic laboratory notebook records; and complete discovery trajectories that connect hypotheses, actions, observations, and revisions, including failed attempts. Supervised fine-tuning supplies worked examples of reasoning, protocol design, and analysis. Reinforcement learning then trains decision-making against research outcomes.

Numbers the company published—and how to read them

On FrontierScience-Research (FSR), Deep Principle reports that Hydrogen-1.0 outperforms GPT-6-Astra and improves over the GLM-5.3 starting checkpoint on GPQA-Diamond, roughly matching the open frontier model Kimi-K3 in its own evaluations. Those figures are self-reported; no independent third-party reproduction is cited in the release.

Internal research-task comparisons against GPT-6-Astra at maximum thinking show Hydrogen-1.0 ahead by 5.13 percentage points on scientific computing and 4.76 points on one-step forward reaction prediction, while trailing Astra by 3.33 points on NMR structure elucidation. The company warns that absolute scores differ by task difficulty and should not be read as a single capability ranking.

A more campaign-shaped demo tracks transition-metal complex design for larger HOMO–LUMO gaps. Each model starts from the same initial population and recommends new candidates over five rounds, receiving only the results of its own picks. By round five, Deep Principle reports mean best-so-far gaps of 4.318 eV for Hydrogen-1.0, versus 3.689 eV for GPT-6-Astra and 3.353 eV for Kimi-K3. Again, these are company-run comparisons.

From a research intent to a compiled lab protocol

Beyond multiple-choice science, the blog walks through practical scenarios: writing and repairing a polyethylene melt workflow in GROMACS; growing an Ireland–Claisen rearrangement network of 27 structures across 30 plausible reactions; and narrowing seven monoformylation candidates of 2-bromonaphthalene against a supplied ¹H NMR trace.

A lab-compilation example is meant to show the gap between a nice protocol and something a particular lab can actually run. Hydrogen-1.0 writes a JSON protocol for a 12-vial Suzuki coupling study. A virtual rehearsal catches an attempt to aspirate without a fresh tip; the model reorders the actions and successfully compiles the revised protocol before execution. Local checks cover station capabilities, dependencies, and action prerequisites—framing that Deep Principle ties to safety concerns about greater AI autonomy in wet labs.

Experiment-planning rubrics on 117 shared questions are reported against Kimi-K3; GPT-6-Astra and Opus-5.5 were omitted from that chart because they rejected more than half of the requests in Deep Principle’s setup.

Serving the weights and what is still closed

Deployment was tested with SGLang on eight GPUs. The card’s launch command uses tensor parallelism of eight, trust-remote-code, GLM-family reasoning and tool-call parsers, and per-request reasoning_effort through chat-template kwargs. Operators must keep generation_config.json so end-of-turn tokens stop generation correctly. Available context and output budgets still depend on the runtime and hardware.

Access routes listed by Deep Principle are Hugging Face and ModelScope downloads, the NeuroSpark API, and Mira, the company’s AI Scientist product. Independently authored Deep Principle code and documentation are Apache-2.0; the BF16 model weights themselves are MIT. The company credits UniPat and Mind Lab on data, training pipeline, and infrastructure work, and NeuroSpark on training infrastructure.

Hydrogen-1.0 does not claim to replace peer review or wet-lab validation. The card states that outputs should be checked against assumptions, constraints, and evidence, and should not be treated as verified experimental findings. What the release does offer is a large open scientific MoE with a long context, concrete lab-oriented demos, and a transparent admission that the recipes behind the post-train stay private.