Unverified50% confidenceFactExact time
Hydrogen-1.0 was not pretrained from scratch but post-trained on Z.ai's GLM-5.3.
1
Sources
50%
Confidence
Long-term
Relevance
11/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe specific composition and cleaning methods of Hydrogen-1.0's training data are not disclosed, which affects outside judgment of the model's capability boundaries.75% similarUnverifiedISO 8573-1 Class 0 does not set a fixed threshold, it is agreed by both supplier and buyer in the contract, must be stricter than Class 1, and must be supported by third-party test reports; Class 1 specifies a total oil content of no more than 0.100 mg/m³.53% similarUnverifiedNo doubts about the GLM-5.3 model weights or training data appeared in the discussion; the focus remained on the inference service's cache management.52% similarUnverifiedMistral Large 4 was trained entirely from scratch in Europe, rather than being post-trained on top of a Chinese open-source model.52% similarUnverifiedFireworks did not provide specific parameter scale, training costs, or real-world performance results for GLM-5.3 on programming and agent tasks in this announcement.51% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1128049API
curl https://kongchang.com/api/v1/knowledge/claims/1128049MCP
get_claim(id=1128049)