AI news
Hydrogen-1.0: A GLM-5.3 Science Model Went MIT-Open Four Days Ago and Nobody Noticed
Short answer
- Deep Principle's Hydrogen-1.0 — a frontier-science model post-trained on top of GLM-5.3 — has had its weights on Hugging Face since Oct 31, under an MIT license, and almost nobody in the English-speaking world has noticed: the three repos still sit at double/triple-digit downloads as of Nov 4.
- What you can actually pull today: the MXFP4 and NVFP4 quantized variants are fully open; the full bf16 repo (~753B safetensors params, same size class as GLM-5.3) is gated:auto — log in, auto-approve, download.
- Our read: this is a specialist research model, not an API alternative. If you're not running math/physics/engineering workloads, skip it. If you are, MIT on a GLM-5.3-class base with 4-bit quants ready is the most permissive science-model package on the board right now — and the self-host bar (~753B bf16, or a serious 4-bit rig) is exactly why API access still matters.
Deep Principle (深度原理) is a Chinese AI-for-science startup, and Hydrogen-1.0 is their post-training of zai-org/GLM-5.3 targeted at scientific workloads — math construction, physics modeling, earth/environment systems, engineering tasks, per the company's own release materials. The interesting licensing move: everything ships MIT, with no MaaS carve-outs or usage restrictions attached. Compare that with the fine-print-heavy licenses some frontier labs attach to their open weights — if you want to finetune Hydrogen for your own materials-simulation pipeline and sell access to the result, the license does not care.
The numbers we could verify first-hand
- 753,329,940,480 safetensors parameters on the base repo — same size class as its GLM-5.3 foundation.
- Repos created 2026-10-31, last modified Nov 2; quantized variants gated: false, base gated: auto.
- Downloads at check time: base 56, MXFP4 1,047, NVFP4 572 — effectively an undiscovered release outside Chinese-language coverage.
- License tag chain confirms
base_model:finetune:zai-org/GLM-5.3andlicense:miton all three repos.
What we can't verify (and won't pretend to)
The benchmark claims — beating GPT-6 Astra and Claude Opus 4.8 on certain science benchmarks — are vendor-reported on their own evaluation suites. We haven't independently run Hydrogen, and science-benchmark contamination is notoriously hard to audit from the outside. Treat the capability claims as "promising, unverified"; treat the licensing and repo facts above as checked.
Practical decision path for science teams: if your workload fits 1M-token-class math/physics prompts and you already run GLM-5.3 locally, swapping the post-trained Hydrogen weights is a drop-in experiment worth a week. If you were hoping for a cheap API — there isn't a public one at time of writing; this is a weights-only release. And if you just want MIT-licensed Chinese frontier weights for general coding or chat, you're better served by the GLM-5.3 base itself or Qwen's ecosystem, which we covered at Wuzhen earlier this week.
Repo facts (params, licenses, gating, download counts, creation dates) verified against the Hugging Face API via mirror on 2026-11-04. Capability claims are vendor-reported and unverified by us. This is a weights-only release as of publication; if Deep Principle opens an API later, that changes the calculus — check before deciding.
Sources: deep-principle/Hydrogen-1.0 model API (checked 2026-11-04), MXFP4 variant listing, 53AI — Deep Principle releases Hydrogen-1.0 with open weights