Deep Principle Opens Hydrogen-1.0 on GLM-5.3, Open Weights Still Need a Runtime

Deep Principle’s official blog lists Hydrogen-1.0 under the date 2026-11-02 (Nov 2, 2026 on the Chinese index). The English landing page is introducing-hydrogen-1.0; the Chinese twin is introducing-hydrogen-1-0-zh. The drop is a third-party scientific post-training release on zai-org/GLM-5.3, aimed at chemistry, materials science, and life sciences—not a new Zhipu / Z.AI flagship product. Product ownership sits with Deep Principle.

Open weights under an MIT card answer a licensing question. They do not answer whether a compatible runtime is already standing. Hugging Face and ModelScope pages are for pulling weights. The Usage examples on the model cards call an already running compatible inference service; they do not download or load weights, and they do not mean either hub ships a ready hosted endpoint. Full training datasets and complete recipes are not released with the weights. The rest of this guide maps those boundaries against the blog and the three Hugging Face model cards. We have not deployed Hydrogen-1.0 end-to-end, have not re-run the blog’s benchmark tables, and have not measured throughput, memory, or scientific-task success on local stacks.

Related reading: Atria Dawn Formal Weights Are Out: 512K Context on Hugging Face, MiniMax M3.1 Flash Open Weights Are on Hugging Face, FlashX Legal Is on OpenRouter Now — Formal ID and Live Prices, and Reproducing Tencent Hyra HyInv and HyPLM-Flash: Gaps That Remain. Those pieces cover other open-weights or GLM-5.3 hosting lines; this article stays on Deep Principle’s Hydrogen-1.0 weights, deploy notes, and openness limits.

Who shipped it, and what the base is

Publisher on the cards: Deep Principle. Base model tag and finetune relation: zai-org/GLM-5.3. The technical blog frames Hydrogen-1.0 as post-training on domain data plus laboratory records and discovery trajectories from Deep Principle’s own labs. That framing is third-party scientific post-training on GLM-5.3, not a Zhipu flagship launch under a new name.

Independent Deep Principle code and documentation elsewhere can carry Apache-2.0; third-party files keep their own licenses. For the weight artifacts discussed here, the cards state MIT for the model weights—retain copyright and permission notices on reuse. The whole tree is not “one license only.”

BF16 main repo: size, gate, context claims

Main repository: deep-principle/Hydrogen-1.0.

Hugging Face API:

Model-card table lines also list context length 1M tokens and maximum output length 128K tokens. Available context and output budgets still depend on the inference framework, service configuration, and runtime resources. The 1M figure is a model-card claim, not proof that every machine boots to a full 1M window out of the box.

Training overview points back to the technical blog. The same cards state that full datasets and complete training recipes are not released with the weights.

Deploy notes on the BF16 card

Visible Deployment section on the main card: the team reports testing with SGLang on 8 GPUs. Keep generation_config.json—it lists end-of-turn tokens; without it, generation does not stop at the end of the answer.

Documented launch shape (IDs and flags as on the card):

python -m sglang.launch_server \
  --model-path deep-principle/Hydrogen-1.0 \
  --served-model-name hydrogen-1.0 \
  --tp 8 --trust-remote-code \
  --reasoning-parser glm45 --tool-call-parser glm47 \
  --host 0.0.0.0 --port 8001

Flags worth keeping in a runbook:

• Tensor parallel --tp 8
• --trust-remote-code
• --reasoning-parser glm45
• --tool-call-parser glm47
• Served API model ID hydrogen-1.0

The Usage example that follows is an OpenAI-compatible client against MODEL_BASE_URL—it assumes the service is already up. That is a call pattern, not a “clone the hub and chat” path, and not proof that Hugging Face or ModelScope host a public inference endpoint for this ID.

Quantized sibling repos: NVFP4 and MXFP4

BF16 is not the only artifact. Two independent quantized repositories sit beside the main repo. Both cards: MIT, base deep-principle/Hydrogen-1.0, and training overview again notes that full datasets / recipes are not shipped with the weights. Quantized outputs may differ from BF16; validate for the intended task before treating results as interchangeable. That caveat is on the cards—we have not measured the delta.

NVFP4 — deep-principle/Hydrogen-1.0-NVFP4

Example launch on the NVFP4 card uses SGLang with --quantization modelopt_fp4 --kv-cache-dtype fp8_e4m3, the same parser flags as BF16, and --context-length 262144. That command caps deployed context at 262144 tokens—not a claim that every setup runs the full 1M. Keep hf_quant_config.json and config.json with the checkpoint.

MXFP4 — deep-principle/Hydrogen-1.0-MXFP4

The MXFP4 card also records a deployment-side concurrency observation under the team’s TP4 test: 128 concurrent long-reasoning requests filled the KV pool and triggered eviction with a sharp quality drop; at 48 concurrent requests the pool stayed below 75% and quality looked normal. Treat that as a reported observation for that setup, not a universal concurrency ceiling—and we have not retested it.

Four official access routes

Blog Availability lists four routes:

1. Hugging Face — download open weights
2. ModelScope — download open weights (ModelScope page for deep-principle/Hydrogen-1.0 returned HTTP 200 when checked)
3. NeuroSpark — access through an API
4. Mira — use Hydrogen-1.0 inside Deep Principle’s AI Scientist product

Routes 1–2 are weight hubs. Routes 3–4 are product / API surfaces. None of those bullets turn the HF or ModelScope model page into “download equals free managed inference.” NeuroSpark and Mira are not unlimited free hosting either—the blog only names the access paths.

Blog eval numbers

The official blog self-reports comparisons on FrontierScience-Research (FSR), GPQA Diamond, and internal research scenarios (scientific computing, forward reaction prediction, NMR elucidation, experiment planning, materials discovery trajectories), including figures versus GPT-6-Astra, Kimi-K3, and the GLM-5.3 starting checkpoint. Those numbers are blog self-reports. We have not re-run those scores. We have not measured deploy speed, VRAM, throughput, or scientific-task success rates ourselves. Blog figures stay under that label—not DigIntoAI measurements.

Operator checklist

1. Attribution. Publisher = Deep Principle; base = zai-org/GLM-5.3; not a Zhipu flagship rename.
2. Clock. Blog index date 2026-11-02; HF main lastModified 17:00:24 Asia/Shanghai on that calendar day (09:00:24 UTC).
3. License vs gate. Weights MIT; main repo gated=auto; NVFP4 / MXFP4 APIs gated=false when checked.
4. Size. BF16 ≈ 753B (API total 753329940480).
5. Context / output. Card: 1M / 128K—usable length follows runtime and memory; NVFP4 example caps context at 262144.
6. BF16 runtime. SGLang, --tp 8, keep generation_config.json, trust-remote-code, reasoning-parser glm45, tool-call-parser glm47.
7. Quant path. Pick NVFP4 (Blackwell / ModelOpt) or MXFP4 (Instinct / Quark / ATOM) as a separate repo; expect possible output drift vs BF16.
8. Usage example ≠ hosting. Client samples talk to a local compatible server; hubs do not become managed endpoints by that alone.
9. Training gap. Public overview on the blog ≠ full datasets / complete recipes in the weight repos.
10. Scores. Blog eval tables = self-reported; we have not retested them.

Bottom line

On 2026-11-02, Deep Principle published Hydrogen-1.0 as open-weight scientific post-training on GLM-5.3, with MIT BF16 weights on Hugging Face / ModelScope plus separate NVFP4 and MXFP4 repos, and named NeuroSpark / Mira as other access routes. That is a clear weights + deploy-notes story with explicit openness limits: gated main repo, no full data/recipe dump, and hub pages that are not “download and chat.” Standing up SGLang or ATOM still requires matching hardware and a working runtime—open permission is not the same as ready to run. Live deploy notes can go through the site-owner contact in the footer.