Attributed compile + one small offline run on my box

Primary source: Introducing Hydrogen-1.0 by Deep Principle, published 2026-11-02.

Weights: deep-principle/Hydrogen-1.0 (BF16, gated: auto, MIT, base zai-org/GLM-5.3) · Hydrogen-1.0-NVFP4 (ungated) · Hydrogen-1.0-MXFP4 (ungated). Local USAGE / NVFP4 card copies: /workspace/reputation/runs/32-hydrogen-1.0/.

Secondary compile: forkalt — Hydrogen-1.0 (hardware / usedStorage notes; not a quality bench).

Tracker mention: The Agentic Weekly — models listed Hydrogen-1.0 among newest verified entries (Nov 2).

Benchmarks below are vendor-reported (blog prose / figures). This box did not download weights, load a checkpoint, or call chat completions. Offline preflight = public HTTP scrapes + HF API JSON only.

Deep Principle shipped Hydrogen-1.0 on 2026-11-02: an open-weight scientific MoE post-trained from zai-org/GLM-5.3, aimed at chemistry, materials science, and life sciences. The HF lead card was created around 2026-10-31; the useful builder story is not “another 万亿 science pitch.” It is four concrete checks: (1) MIT weights with a gated BF16 path and ungated NVFP4 / MXFP4 siblings, (2) official SGLang / ATOM recipes that actually cap deployable context, (3) a client surface that is OpenAI-compat but fussy about reasoning_effort placement, and (4) vendor claims that Hydrogen beats GPT-6-Astra on FrontierScience-Research (FSR) and improves over GLM-5.3 on GPQA-Diamonds — which is a reason to run your golden science eval, not a reason to paste their PNG into a go/no-go slide.

Four access routes are named on the blog: Hugging Face, ModelScope, NeuroSpark API, and Mira (Deep Principle’s AI Scientist). This box opened HF pages + APIs and the launch blog; ModelScope / NeuroSpark / Mira stay blog labels here — we did not authenticate or call them.

What it is (short)

Hydrogen-1.0 is a text-only (glm_moe_dsa / GlmMoeDsaForCausalLM) scientific problem-solving and research-assistance model. The chat-template language on the card reminds you there is no multimodal input. Domains called out: chemistry, materials, life sciences. Training story (vendor): SFT + RL on domain knowledge, research scenarios, laboratory records, and discovery trajectories — full datasets and recipes are not released with the weights. Card / API context claim: 1M tokens in, 128K max out. API config we saw via preflight needles: num_experts_per_tok 8. HF reports ~753B parameters; safetensors total 753,329,940,480 (same total forkalt cites vs the GLM-5.3 base).

Access matrix

Route What we saw Gate / license note
HF BF16 deep-principle/Hydrogen-1.0 Card + API HTTP 200 gated: auto; license: mit; accept the HF form before shards download
HF NVFP4 Card + API HTTP 200 ungated; MIT LICENSE © 2026 Deep Principle and Z.AI; ModelOpt / Blackwell / 8× B300 recipe
HF MXFP4 Card + API HTTP 200 ungated; Quark / AMD Instinct / MI350X + ATOM recipe
ModelScope / NeuroSpark / Mira Named on blog only Not opened in this preflight — treat as blog labels

Our 2026-11-03 ~22:04 Asia/Shanghai HF API preflight (not forkalt’s earlier snapshot): BF16 likes 200 / downloads 56; NVFP4 likes 7 / downloads 572; MXFP4 likes 8 / downloads 1047. Quant siblings already moving downloads while the gated BF16 likes lead — classic “try the smaller disk first” pattern.

Hardware / size (cite sources)

From HF API + forkalt usedStorage notes (we did not huggingface-cli download):

  • Params: ~753B; safetensors total 753,329,940,480.
  • BF16 usedStorage (forkalt): ~1.37 TiB across 303 shards.
  • NVFP4 usedStorage (forkalt): ~432 GiB.
  • MXFP4 usedStorage (forkalt): ~394 GiB.

Peak VRAM and tokens/s are not independently measured here. Plan rack space from usedStorage + the team’s tp 8 (BF16/NVFP4 SGLang) or tp 4 (MXFP4 ATOM) recipes, not from the 1M marketing number alone.

Deploy traps (the builder core)

Keep the checkpoint’s tokenizer, chat template, weight index, quant configs, and especially generation_config.json. Cards warn that stop / end-of-turn tokens live there; drop it and generation may not stop at the answer.

BF16 SGLang example (from card / forkalt copy — not run here):

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 SGLang example (team-tested language: 8× B300) is where the context cliff shows up:

python -m sglang.launch_server \ --model-path deep-principle/Hydrogen-1.0-NVFP4 \ --served-model-name hydrogen-1.0 \ --tp 8 --trust-remote-code \ --quantization modelopt_fp4 --kv-cache-dtype fp8_e4m3 \ --reasoning-parser glm45 --tool-call-parser glm47 \ --context-length 262144 --mem-fraction-static 0.85 \ --chunked-prefill-size 16384 \ --host 0.0.0.0 --port 8001

Card context is 1M; this recipe sets --context-length 262144. That is the cliff: your prompt + generation must fit the deployed budget, not the brochure. Raising it needs compatible runtime support and enough memory — do not assume 1M is live because the model card says 1M.

Do not mix BF16 / NVFP4 / MXFP4 files (configs, indexes, shards). USAGE.md is explicit.

Client: OpenAI-compat against a running service (SERVED_MODEL_NAME=hydrogen-1.0). For reasoning effort, pick one convention your server actually forwards:

  • extra_body={"chat_template_kwargs": {"reasoning_effort": "max"}}, or
  • top-level reasoning_effort="max"

Do not stack both by default. Card demos use max and max_tokens=4096 as an illustrative budget, not the 128K max-out claim. Check finish_reason=length (and empty final content) before you treat a truncated stream as a scientific answer.

Vendor benches (label as vendor)

From the 2026-11-02 blog prose only. FSR / GPQA numeric table rows were not OCR’d from the PNG figures — do not invent scores here.

  • Vendor: Hydrogen-1.0 outperforms GPT-6-Astra on FrontierScience-Research (FSR) and improves over GLM-5.3 on GPQA-Diamonds, roughly matching Kimi-K3 in their evaluations.
  • Internal task deltas (vendor prose vs Astra at max thinking): scientific computing +5.13 pp, forward reactions +4.76 pp; Astra still ahead on NMR structure elucidation by 3.33 pp.
  • TMC discovery (vendor): by round five, mean best-so-far HOMO–LUMO gap 4.318 eV (Hydrogen) vs 3.689 eV (Astra) / 3.353 eV (Kimi-K3), same initial population, independent feedback histories.

Useful as a map of where they want you to look. Useless as a substitute for your own tickets, tools, and lab constraints. Same series rule: vendor framing ≠ merge into prod.

Checklist before you trust it for science work

  1. Freeze a golden eval on your scientific tickets (derivations, reaction / NMR cases, protocol drafting, tool-using MD or ELN loops you actually run). FSR / GPQA are someone else’s harness.
  2. Gate accept on BF16 if you need unquantized weights; or pick the ungated NVFP4 / MXFP4 path that matches your GPUs.
  3. Pick one quant path and keep its files together — no BF16/NVFP4/MXFP4 mashups.
  4. Budget context from the launch flags, not the 1M card line. NVFP4 recipe = 262K until you prove otherwise.
  5. Pin reasoning_effort placement (template kwargs or top-level) and pin served-model-name hydrogen-1.0.
  6. Treat outputs as drafts. Card language: assess against assumptions, constraints, and evidence; do not treat model text as verified experimental findings or as grounds for health/safety decisions.
  7. Watch truncation: demo max_tokens=4096 is not a research budget; inspect finish_reason.
  8. Do not conflate blog FSR/GPQA claims, internal pp deltas, TMC eV gaps, and any third-party index you have not opened yourself.

What I actually ran (offline)

/workspace/reputation/runs/32-hydrogen-1.0/ — 2026-11-03 22:04 Asia/Shanghai (CST):

NO API KEY / NO WEIGHT DOWNLOAD / NO CHAT COMPLETION / NO SGLang HTTP 200: blog.deepprinciple.com/introducing-hydrogen-1.0/ HTTP 200: huggingface.co/deep-principle/Hydrogen-1.0 (+ API JSON) HTTP 200: huggingface.co/deep-principle/Hydrogen-1.0-NVFP4 (+ API JSON) HTTP 200: huggingface.co/deep-principle/Hydrogen-1.0-MXFP4 (+ API JSON) HTTP 200: forkalt.com/models/hydrogen-1-0/ HTTP 200: theagenticweekly.com/news/models HF API (this preflight): BF16 likes 200 dl 56; NVFP4 likes 7 dl 572; MXFP4 likes 8 dl 1047 Local copies: USAGE.md, NVFP4-README.md, MXFP4-README.md, LICENSE (MIT © 2026 Deep Principle / Z.AI) NOT tested: weight download; SGLang/ATOM serve; chat completion; ModelScope; NeuroSpark API; Mira; FSR/GPQA re-score; OCR of blog PNG figures; VRAM measure

When Hydrogen-1.0 might become a science default

Only after a callable path and a golden eval win on your problems:

  1. You need an open MIT science MoE and can accept gate + disk (BF16) or a Blackwell / Instinct quant path.
  2. You have chosen quant + context budget knowingly (especially the 262K NVFP4 cliff).
  3. reasoning_effort and stop tokens are pinned; truncation is monitored.
  4. Vendor FSR / Astra framing is treated as a map, not a certificate.
  5. You can roll back without rewriting the lab toolchain.

If you only have a laptop GPU and no B300 / MI350X class box, this drop is the wrong trigger for local serve — use the blog’s API/product labels only after you open them yourself. If you already run SGLang on eight high-memory GPUs and your failures are chemistry / materials reasoning tickets, Hydrogen-1.0 is a same-week candidate for a side-by-side, not because the Astra FSR sentence is gospel, but because the weights and recipes are concrete enough to instrument.

Hydrogen-1.0’s builder-relevant release is the MIT open-weight science MoE, the gated BF16 vs ungated NVFP4/MXFP4 split, the 262K context cliff on the NVFP4 recipe, and vendor FSR / GPQA / internal deltas labeled as such. Treat those as inputs, run your own eval, and only then change the production model string.

— YongBo Yu · Toronto AI engineer · yongbo-yu.vercel.app · GitHub @YongBoYu1