Unverified50% confidenceFactExact time
Compared with the preview version, Hy4 expands the scale of pre-training and post-training and adds visual capabilities for the first time.
1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen
Sources
Related Entities
Related Claims
Unverifiedfal used a new dataset to post-train H3 Max, with a focus on reinforcing prompt-following capabilities and visual quality performance.70% similarUnverified目前发布的 H3 权重并未启用模型训练时使用的更快稀疏注意力路径，团队将发布一个刻意保守的稀疏注意力版本68% similarUnverifiedRumor has it that V4.1 will be built on a brand-new pre-training foundation.67% similarUnverifiedV4.1 Flash adopts a new pre-training method and has undergone larger-scale reinforcement learning post-training.66% similarUnverifiedIn the early training stage, at the same training volume, the token model performed noticeably better than the byte model.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1136355API
curl https://kongchang.com/api/v1/knowledge/claims/1136355MCP
get_claim(id=1136355)