Deep Principle’s Hydrogen-1.0: test open-weight scientific post-training—not a “superintelligence for science” slogan

The claim worth testing in Hydrogen-1.0 is not a “superintelligence for science” slogan, but whether scientific post-training of an open-weight base on lab trajectories and domain corpora yields stable gains on FSR / GPQA / TMC-style settings versus GLM-5.3 (and the post’s comparison models).

Falsifiable read: if third parties with the open weights cannot reproduce Deep Principle’s reported FSR/GPQA/TMC-style gains, treat the launch mainly as vendor direction and an internal eval dashboard; if independent reproduction holds, raise the weight of lab-trajectory post-training.

Hard facts (Deep Principle official blog): Hydrogen-1.0 comes from post-training GLM-5.3; data spans chemistry, materials, and life sciences plus lab records and agent trajectories; training described as SFT + RL with Megatron and SGLang, CUDA and ROCm. Collaborators named: UniPat, Mind Lab, NeuroSpark.

Product side: open weights; contrasts with proprietary discovery-model narratives it cites (OpenAI Navier–Stokes/Lean example; Anthropic CRISPR-like enzyme claim—reported here as Deep Principle’s citations, not independently verified in this round). Offers ELN + Mira agentic documentation and lab compilation / virtual rehearsal for protocols (Suzuki coupling tip-aspiration example in the post).

Evidence strength: first-party launch plus self-reported benches and internal head-to-heads; workflow examples concrete enough (structure/reaction counts, eV figures) for third parties to design reproduction protocols. But GPT-6-Astra / Kimi-K3 / GLM-5.3 names and scores are Deep Principle’s evaluations; cost axes also apply GLM-5.3 pricing to Hydrogen.

Steelman: writing real lab trajectories into post-training is closer to an experimental loop than general science QA alone; open weights lower the reproduction bar versus closed discovery models. Beyond the steelman: self-boards bias to success paths; trailing Astra by 3.33pp on NMR shows non-dominance; without an independent eval pack, “beats Astra on FSR” remains a pending sentence.

Strongest steelman: if open-weight scientific post-training truly turns lab trajectories into transferable skill, smaller labs could run near-closed discovery workflows locally or via controlled APIs—rather than buying only a chat UI.

Steelman risk: most figures are self-reported; our notes lack a citable exact Hugging Face repo ID; comparison-model versions, prompts, and tool permissions are not fully pinned in this round’s materials; “open weights” still needs actual uploads and licenses.

For scientific-ML and lab-automation readers, watch whether weights are downloadable, whether FSR/GPQA/TMC protocols can be rerun, and whether Mira/ELN trajectories export auditably—not whether “science superintelligence” rhetoric grows louder. For buyers, treat vendor head-to-heads as direction signals and third-party reruns as closing conditions.

Near term, do not read self-reported FSR/GPQA/TMC gains as settled scientific-model SOTA: judge by downloadable open weights, third-party reruns under comparable protocols, and failure disclosure. Missing those three, Hydrogen-1.0 is mainly a product narrative about post-training direction; with them, it can become a reference point for the “lab trajectories → open-weight scientific model” path.

【No.2】