In Hydrogen-1.0 vs GPT-6-Astra, the open-weight model wins most lab-workflow science tasks in Deep Principle's own evaluations, scoring 38.33% against 33.33% on FrontierScience-Research with Astra at maximum thinking. Hydrogen-1.0 is also cheaper per problem on that benchmark, at $0.25 against $0.88. GPT-6-Astra keeps one clear advantage, NMR structure elucidation, where it leads by 3.33 points. The weights went up on Hugging Face on November 2, 2026. Every number here is self-reported by the vendor and has not been independently replicated.
TL;DR
- Hydrogen-1.0 is an MIT-licensed, 753B-parameter mixture-of-experts model that Deep Principle post-trained from Z.ai's GLM-5.3 for chemistry, materials science and life sciences.
- On FrontierScience-Research it scores 38.33% versus 33.33% for GPT-6-Astra (max thinking), per the company's launch post.
- It also leads on scientific computing (76.92% vs 71.79%) and one-step forward reaction prediction (+4.76 points). Astra wins NMR structure elucidation.
- Cost per problem is $0.25 versus $0.88, though the Hydrogen-1.0 figure uses GLM-5.3 pricing as a proxy.
- Weights are on Hugging Face in BF16, NVFP4 and MXFP4. Training data and recipes are not released.
- Treat the results as a strong claim awaiting third-party replication, not a settled ranking.
What is Hydrogen-1.0 and who built it?
Hydrogen-1.0 comes from Deep Principle, a company building AI tools for chemistry and materials research. Founder and CTO Chenru Duan announced it on X on 2 November 2026 and the weights were posted the same day, according to Tech City Authority. A formal press release followed on 6 November.
The model has 753B parameters in a mixture-of-experts design, the same base as GLM-5.3. The change is in the training. Deep Principle used supervised fine-tuning and reinforcement learning on scientific data, electronic lab notebook records, and agent discovery trajectories that include failed attempts and unexpected results, according to its launch post. Partners UniPat and Mind Lab contributed to data, pipeline and infrastructure, with NeuroSpark providing training infrastructure.
Who wins in Hydrogen-1.0 vs GPT-6-Astra on science tasks?
Hydrogen-1.0 wins on four of the five measures Deep Principle reports. Astra wins the fifth. All figures below come from the company's blog post and press release.
| Measure (self-reported) | Hydrogen-1.0 | GPT-6-Astra (max) | Kimi K3 | GLM-5.3 |
|---|---|---|---|---|
| FrontierScience-Research pass rate | 38.33% | 33.33% | 29.72% | 24.44% |
| Cost per problem (output tokens) | $0.2500 | $0.8824 | Not reported | Not reported |
| Average output tokens | 56,826 | 17,649 | Not reported | Not reported |
| Scientific computing | 76.92% | 71.79% | Not reported | Not reported |
| Closed-loop discovery, best HOMO-LUMO gap by round five | 4.318 eV | 3.689 eV | 3.353 eV | Not reported |
GPQA-Diamond is a general graduate-level science test rather than a lab-workflow one. Hydrogen-1.0 reaches 93.0%, an improvement on GLM-5.3 that roughly matches Kimi K3. The sources I reviewed give no GPQA figure for Astra, so that comparison is left open.
For readers who follow open-weight releases generally, our coverage of Kimi's open-weight release gives useful context on the competing model family.
How much cheaper is Hydrogen-1.0 per problem?
On FrontierScience-Research, Hydrogen-1.0 scores 5.00 points higher than Astra at max thinking while the launch post puts the cost gap at $0.8824 minus $0.2500, or $0.6324 per problem, 71.7% of Astra's cost. Both figures include reasoning tokens and value Hydrogen-1.0 at GLM-5.3 pricing as a proxy, because NeuroSpark has not published per-token pricing.
The token counts add a wrinkle. The launch post reports 56,826 average output tokens per problem for Hydrogen-1.0 against 17,649 for Astra, so it talks roughly three times as much while paying for it with a lower per-token price. Dividing cost by tokens implies about $4.40 per million output tokens for Hydrogen-1.0 and about $50 for Astra, my derivation from the published per-problem costs and token counts, not a vendor price list. It shows the saving comes entirely from a lower per-token price, which the company says uses GLM-5.3 pricing as a proxy. NeuroSpark has not published per-token API pricing, so real-world cost is unconfirmed.
Does post-training beat scale for scientific work?
This result suggests that domain post-training can matter more than a larger or more expensive general model, at least for lab-style tasks. Hydrogen-1.0 shares its 753B base with GLM-5.3, which scores 24.44% on the same benchmark. Post-training lifted it to 38.33%, per the same evaluation tables, a gain of 13.89 points with no change in size.
The likely driver is the data. Lab notebooks record what researchers actually tried, including dead ends. Trajectories that include failures teach a model when to stop, repair an error or change approach. The company's polymer demo shows the pattern: the agent built a polyethylene melt in Python, ran GROMACS, detected a topology error, repaired it, restarted, and analysed the corrected 40-nanosecond trajectory.
The caveat is that one benchmark, run by the model's builder, is thin evidence. The company itself says the scales of its internal evaluations should not be read as direct comparisons of model capability.
What about the materials discovery loop?
In a five-round transition-metal-complex design task aiming for larger HOMO-LUMO gaps, the mean best-so-far gap by round five was 4.318 eV for Hydrogen-1.0, 3.689 eV for GPT-6-Astra and 3.353 eV for Kimi K3. That is a gap of 0.629 eV between Hydrogen-1.0 and Astra. A larger gap is not automatically a better material, so this measures optimisation skill rather than usefulness.
What can a lab actually do with the weights?
The model is released under the MIT License, so commercial and academic use are both permitted. The Hugging Face card lists three builds:
- BF16, the unquantised reference weights.
- NVFP4, for NVIDIA Blackwell B200 and B300 GPUs (supercomputing.news).
- MXFP4, for AMD Instinct MI350X and MI355X GPUs.
The card reports a 1M-token context window and 128K maximum output, with deployment tested on SGLang across 8 GPUs using --tp 8. Weights are also on ModelScope, and the model is available through a NeuroSpark API and inside Deep Principle's Mira AI Scientist product. The training datasets and recipes are not part of the release, so you can run and fine-tune the model but not reproduce its training.
Smaller specialist models are a separate story; see our piece on GroundingPi, a 4B grounding model compared with GPT-6-Astra. For Astra's position against other proprietary models, our GPT-6-Astra vs Claude Fable 5.1 verdict covers the general-purpose picture, and the Kimi K3.1 Pro vs Claude Opus 5.5 comparison covers the open-versus-closed question outside science.
Why we picked this matchup
We track which of our comparison topics face no established competitor: within our 72 winnable keywords, 20 carry a difficulty score of zero, meaning no established competitor holds the result set (n=72, DataForSEO difficulty data, measured 2026-11-05). This Hydrogen-1.0 matchup is one of them.
Which should you choose?
Choose Hydrogen-1.0 if you run chemistry or materials workflows, have your own GPUs or want control over data, and can tolerate a vendor-reported evidence base. Choose GPT-6-Astra if your work leans on NMR structure elucidation, or if you want a hosted service with an established support path. Run your own evaluation on a sample of real problems before committing either way.
FAQ
Q: Is Hydrogen-1.0 really better than GPT-6-Astra for science?
A: On FrontierScience-Research and scientific computing, yes, in Deep Principle's own evaluations. GPT-6-Astra wins NMR structure elucidation by 3.33 points. Third-party replication is still pending, so treat the lead as provisional.
Q: Is Hydrogen-1.0 free to use?
A: The weights are free to download and use under the MIT License on Hugging Face and ModelScope. You pay for your own hardware. The NeuroSpark API and the Mira product are separate services, and NeuroSpark has not published per-token pricing.
Q: What hardware do I need to run it?
A: The documented deployment uses SGLang across 8 GPUs. NVFP4 builds target NVIDIA Blackwell B200 and B300, and MXFP4 builds target AMD Instinct MI350X and MI355X. A single consumer GPU is not a realistic target.
Q: What is FrontierScience-Research?
A: It is a benchmark of open-ended, research-style science problems, scored as a pass rate. In this comparison Deep Principle ran it itself, with Hydrogen-1.0 at 38.33% and GPT-6-Astra at 33.33%.
Q: Does 753B parameters mean it needs 753B worth of GPU compute per token?
A: No. It is a mixture-of-experts model, so only a subset of experts is active for each token. All the weights still have to be stored in memory, which is why multi-GPU deployment is required.
Top comments (0)