Deep Principle ships Hydrogen-1.0, a 753B open science model

Deep Principle released Hydrogen-1.0, a 753B-parameter open-weight model for chemistry, materials science and life sciences, built by post-training Z.ai's GLM-5.3 and shipped under the MIT license. The vendor reports a 38.3% pass rate on FrontierScience-Research against 33.3% for OpenAI's GPT-6-Astra.

Key facts

• 753B parameters, mixture-of-experts architecture with 8 experts per token, BF16 safetensors, base model tagged zai-org/GLM-5.3 and license tagged mit on the Hugging Face repo
• Vendor-reported FrontierScience-Research pass rate of 38.3% for Hydrogen-1.0 versus 33.3% for GPT-6-Astra — a Deep Principle self-report, not independently replicated
• Same vendor evaluations put GPT-6-Astra ahead on NMR structure elucidation by 3.33 percentage points at maximum thinking
• Training combined supervised fine-tuning and reinforcement learning on domain data, scientist feedback, electronic lab notebook records, and complete discovery trajectories
• In a 117-question experiment-planning comparison, GPT-6-Astra and Opus-5.5 were omitted because they rejected more than 50% of requests, per Deep Principle's writeup
• Weights download is offered through Hugging Face and ModelScope, with API access through NeuroSpark and product access through Deep Principle's Mira AI Scientist

What happened

Deep Principle, a company running its own research laboratories, introduced Hydrogen-1.0 on its engineering blog as "a scientific model built by post-training GLM-5.3 on domain data from chemistry, materials science, and life sciences, together with records and agent trajectories from our own laboratories." The Hugging Face repository confirms the mechanics: a glm_moe_dsa text-generation architecture with 8 experts per token, BF16 weights, and explicit base_model:zai-org/GLM-5.3 and license:mit tags. Downloads are gated with an approval step — the repo returns "Access restricted" to unauthenticated requests — but there is no license restriction on use once granted.

The model targets research work rather than chat. Deep Principle's demonstrations include writing and debugging a polymer molecular-dynamics workflow in GROMACS, growing an Ireland–Claisen reaction network across 27 candidate structures, elucidating structures from ¹H NMR traces, and writing JSON protocols for a 12-vial Suzuki coupling study that a "lab compilation" layer checks against equipment limits before execution.

The benchmark story has an asterisk the company itself draws. Deep Principle says Hydrogen-1.0 "outperforms GPT-6-Astra on FrontierScience-Research (FSR) and improves over GLM-5.3 on GPQA-Diamonds in our evaluations, roughly matching the open frontier model Kimi-K3." Every number in that sentence is the vendor's own, run on its evaluations with its cost accounting — Hydrogen-1.0 is priced at GLM-5.3 rates in the comparison figures. No independent lab has replicated the results on the released weights, and the company's internal scenario tests (scientific computing, forward reaction prediction, NMR) are likewise self-reported. The most checkable claims are the structural ones: parameter count, architecture, base model and license, all verifiable in the repository metadata.

The training pipeline is unusually well documented. Deep Principle describes supervised fine-tuning on worked examples of reasoning, experiment planning and protocol design, followed by reinforcement learning on research outcomes. Full-model MoE post-training ran on infrastructure built from Megatron and SGLang spanning NVIDIA CUDA and AMD ROCm, with FP32 master weights, disk-offloaded optimizer momentum, and a bidirectional adapter between training checkpoints and FP8 inference weights. Resumable rollouts let reinforcement-learning responses span multiple policy versions, with per-token behavior log-probabilities retained for importance correction.

Why it matters

The interesting signal here is not the benchmark claim — it is the shape of the release. A domain team took a frontier open base model, added proprietary laboratory data that no API seller has, and shipped the result under a license that lets any lab inspect, adapt and redeploy it. That is the vertical-fine-tune playbook working at 753B scale, and it is the quiet workhorse scenario for private inference: a research organization can run a domain-specialized model against its own unpublished data without sending that data to a hosted endpoint.

The cost dimension matters for builders. Deep Principle's comparison figures fold API cost into every chart, using GLM-5.3 pricing for Hydrogen-1.0. Whatever the scores are worth, the positioning is explicit: scientific capability per output-token dollar, aimed at labs that would otherwise rent a closed frontier API. That framing also gives the model a natural contrast with trusted-hardware privacy plays like OpenAI's Private Inference API, which we covered when it launched in October — confidential compute inside a vendor's boundary, versus weights you hold yourself.

There is also a counterpoint worth stating plainly. A model post-trained on one company's laboratory records inherits that company's instrument fleet, protocol style and research priorities. Whether those gains transfer to a lab with different equipment and different chemistry is exactly the kind of question independent replication answers, and nobody has published one yet.

Background

Vertical fine-tunes of open foundations have become a reliable release pattern on this beat. Recursive's Flint Q1 took an open base and specialized it for coding; we covered its benchmark claims with the same vendor-reported caveat. DeepSeek's V4.1 Flash update in October showed how quickly a base-model patch can ripple through the provider ecosystem. Hydrogen-1.0 is the life-sciences expression of the same dynamic, and at 753B it is the largest example we have covered — a size that puts self-hosting out of reach for most labs and makes the hosted-provider and marketplace listings the practical access route for all but the best-equipped.

The release also lands in a week when training-data provenance is under scrutiny: The Information reported — and TNW relayed without independent verification — that DeepSeek and Moonshot may have distilled from OpenAI's GPT-6 Astra outputs. Hydrogen-1.0's provenance is unusually auditable by comparison: the base model is tagged in the repo metadata, and the training blog post documents its data sources in detail. Provenance is becoming a differentiator, not a footnote.

For readers new to the space, our open-weights releases coverage tracks these drops and the provider economics around them; the inference-provider margin debate currently running on X is a reminder that hosting a 753B MoE is a serious capital commitment, whatever the license says.

What's next

Three things to watch. First, independent replication: when a third party runs FrontierScience-Research and the internal scenario benchmarks against the released checkpoints, the vendor's numbers either harden into results or soften into marketing. Second, the gate: the Hugging Face approval step will show whether adoption is broad or concentrated among institutional users. Third, the listing question — whether hosted providers and decentralized inference marketplaces add a 753B science-specialist MoE, and at what price, will tell you how much demand the vertical actually has. Deep Principle's own NeuroSpark API and Mira product are the reference implementations to check pricing against.