Hydrogen-1.0: Specs, License & How to Run

Hydrogen-1.0 is a text-only scientific MoE from Deep Principle, post-trained from GLM-5.3. Official hosts we opened on 2026-11-03: the gated lead card deep-principle/Hydrogen-1.0 (created 2026-10-31T15:54:37Z; last modified 2026-11-02T09:00:24Z; 136 likes / 0 downloads), ungated quantized siblings Hydrogen-1.0-NVFP4 (3 likes / 96 downloads) and Hydrogen-1.0-MXFP4 (3 likes / 108 downloads), the MIT LICENSE on NVFP4, the launch post Introducing Hydrogen-1.0 (2 November 2026), and the named base zai-org/GLM-5.3. API gated on the BF16 card was auto. Hugging Face pipeline_tag is text-generation; model_type is glm_moe_dsa.

This page hangs off ChatGPT alternatives and closed LLM APIs. That mapping is editorial. We have no sourced quality comparison against ChatGPT. Head-to-head facts versus the open 2.8T card the blog names: Hydrogen-1.0 vs Kimi K3. Same-site GLM-5 family card: GLM-5.3-Flash — a different MIT 320B/18B multimodal SKU, not the GLM-5.3 base. There is no GLM-5.3 identification page here.

The card calls Hydrogen a scientific problem-solving and research-assistance model for chemistry, materials science, and life sciences. Outputs “should be assessed against the assumptions, constraints, and evidence relevant to each task.” Those sentences stay vendor. The chat template we opened on the BF16 API config reminds that the model has no multimodal input.

What we verified (2026-11-03)

Field	Value	Source
Host (BF16)	huggingface.co/deep-principle/Hydrogen-1.0 (gated: auto)	HF API
Created / last modified	2026-10-31T15:54:37Z / 2026-11-02T09:00:24Z	same
Likes / downloads (BF16)	136 / 0	same
License tag (BF16)	mit	HF cardData
LICENSE file we opened	MIT on the NVFP4 sibling (© 2026 Deep Principle; © 2026 Z.AI)	NVFP4 LICENSE
Pipeline tag	text-generation	HF API
Architecture	glm_moe_dsa / GlmMoeDsaForCausalLM	card + API config
Base model	zai-org/GLM-5.3	card YAML / HF tags
Params (HF)	753B; safetensors 753,329,940,480	HF API / UI
Experts named	num_experts_per_tok 8	API config
usedStorage (BF16)	1,506,692,259,475 bytes (~1.37 TiB); 303 shards	HF API
Context / max output	1M tokens / 128K tokens	card table
NVFP4	ungated; 3 likes / 96 downloads; usedStorage ~432 GiB	HF API
MXFP4	ungated; 3 likes / 108 downloads; usedStorage ~394 GiB	HF API

HF tags on the lead card include base_model:zai-org/GLM-5.3 and base_model:finetune:zai-org/GLM-5.3. An active-parameter count is not printed on the Hydrogen card we opened; we do not invent one.

The 2 November 2026 blog names four access routes: Hugging Face weights, ModelScope weights, a NeuroSpark API, and Mira (Deep Principle’s AI Scientist product). We opened Hugging Face and deepprinciple.com (Agent Mira copy). We did not open a working NeuroSpark or ModelScope host on this check, so those names stay blog labels.

How to run (from official cards, not tested)

BF16 card (gated): tested with SGLang on 8 GPUs. Keep generation_config.json — the card says it lists end-of-turn tokens, and without it generation does not stop at the end of the answer. Copied launch names only:

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

NVFP4 card: quantized with NVIDIA ModelOpt; team-tested SGLang on 8× B300. The example we copied caps deployed context at 262,144 tokens, not the card’s 1M, and names --quantization modelopt_fp4 --kv-cache-dtype fp8_e4m3. MXFP4 card names a compatible AMD Instinct runtime, team-tested on MI350X. forkalt has not run those commands.

Client examples on the cards call an already-running OpenAI-compatible service with chat_template_kwargs.reasoning_effort set to max. That is a vendor recipe.

License

LICENSE on the ungated NVFP4 sibling is the MIT License, copyright 2026 Deep Principle and 2026 Z.AI. Lead-card cardData.license is mit. The BF16 README we could open (the gated HTML card) says the BF16 weights are MIT and that independently authored Deep Principle code and documentation are Apache-2.0. We did not fetch the gated LICENSE blob. Not legal advice.

What we did not invent

• Numeric FrontierScience-Research / GPQA-Diamond scores stay on the launch blog figure. We do not OCR that PNG.
• “Outperforms GPT-6-Astra” and “roughly matching Kimi-K3” stay vendor.
• Active-parameter count, peak VRAM, and tokens/s stay unprinted / vendor.
• We do not rank Hydrogen against Kimi K3, GLM-5.3-Flash, Sage-A1.5, or ChatGPT.

Related on this site: Kimi K3 — Hydrogen-1.0 vs Kimi K3 — GLM-5.3-Flash (multimodal Flash SKU, not the GLM-5.3 base) — Sage-A1.5 (35B science/agent VLM) — Carbon-X (DNA+text; not clinically validated) — DeepSeek-V4.1-Flash. Serving: vLLM. Catalog: home.