Reported News
Hydrogen-1.0: Deep Principle's 753B open-weight science model, post-trained from GLM-5.3 under MIT
Hydrogen-1.0 is an open-weight scientific model from Deep Principle, which says it was built "by post-training GLM-5.3 on domain data from chemistry, materials science, and life sciences". Its Hugging Face listing is MIT-licensed and reports about 753 billion parameters; access requires accepting the gated repo's terms.
- Changes reported
- 6
- listed below, each from a source we read
- We tested
- Not tested yet
- reported from sources; we did not run it ourselves
- Gaps we flag
- 5
- things we could not verify
- Sources
- 2
- 2 primary, linked below
What changed
| # | What the source says | Status |
|---|---|---|
| 01 | Deep Principle says Hydrogen-1.0 is GLM-5.3 post-trained on chemistry, materials science and life-sciences data plus records and agent trajectories from its own laboratories. | Reported |
| 02 | The Hugging Face repo lists an MIT license, GLM-5.3 as the base model and 753,329,940,480 parameters in its safetensors weights; downloads are gated behind accepting the repo's terms. | Reported |
| 03 | In Deep Principle's own evaluations it beats GPT-6-Astra on FrontierScience-Research and roughly matches Kimi-K3 on GPQA-Diamond; against Astra at maximum thinking it leads by 5.13 and 4.76 points on scientific computing and forward reactions but trails by 3.33 on NMR structure elucidation. | Reported |
| 04 | On a five-round transition metal complex design test, the team reports a mean best-so-far HOMO-LUMO gap of 4.318 eV for Hydrogen-1.0, versus 3.689 eV for GPT-6-Astra and 3.353 eV for Kimi-K3. | Reported |
| 05 | Training combines supervised fine-tuning with reinforcement learning on research outcomes, built on Megatron and SGLang across both NVIDIA CUDA and AMD ROCm. | Reported |
| 06 | Four access routes are listed: weights on Hugging Face and ModelScope, an API through NeuroSpark, and the company's Mira AI Scientist app. | Reported |
reported by the sourcemeasured or run by usnot verified
Primary source: Deep Principle blog: Introducing Hydrogen-1.0. We are the fast, independent index; the vendor’s own page is the source of record for this release.
Who it matters for
Research teams in chemistry, materials science and life sciences who want a self-hostable model they can inspect and fine-tune, and builders of lab or simulation agents who need protocol writing and experiment planning. General coding or chat users can ignore it: every benchmark in the announcement is scientific, and all of them are run by the vendor.
At roughly 753 billion parameters this is a multi-GPU deployment; the hosted API route is the practical way to try it without that hardware.
In the source’s words
“Hydrogen-1.0 is available through four access routes”
“Astra at maximum thinking still scores higher on NMR structure elucidation by 3.33 points.”
What we have not verified
- We did not run this release ourselves.
- All benchmark and discovery results are Deep Principle's own evaluations; several use internal test sets that have not been published.
- The Hugging Face model card is gated, so we read the repo metadata (license, base model, parameter count) but not the card text; context length and hardware requirements reported in press coverage are not confirmed here.
- We did not check NeuroSpark API pricing or availability.
- The announcement page carries no date, so the source publish time is left blank; the Hugging Face repo was last modified 2026-11-02 09:00 UTC.
This brief is Reported: written from the sources below, not from running the release ourselves.
Sources
- Deep Principle blog: Introducing Hydrogen-1.0primary source · fetched 2026-11-02
- Hugging Face: deep-principle/Hydrogen-1.0 (repo metadata)primary source · fetched 2026-11-02
Related
Spotted an error? See our corrections policy or email the address on the contact page. Markdown version for agents: /news/deep-principle-hydrogen-1-0-open-weight-science-model.md
Want these briefs in your Google Top Stories? Add Agentic Weekly as a preferred source on Google.
Newsletter
The Weekly Verdict. Free, every Monday.
A free email every Monday: the week's few big AI releases, then 5 short use cases for the new models, each with a prompt you can paste.
Free. Double opt-in. One click to unsubscribe.