← Back to live feed1 story
The Mixture-of-Experts system features 49 billion active parameters and a 1 million token context window for productivity and office workflows. Tencent released the Hy4 Preview under an Apache 2.0 license to compete in the open source frontier of coding and scientific research.
The model scored 1,633 points in the Code Arena WebDev AutoEval, ranking approximately 5th overall and 3rd among open weights models. This performance exceeds the previous Hy3 version, which ranked 31st overall, and includes agentic capabilities that coordinate multiple Codex sessions in parallel to adjust research directions.
Sign in to suggest edits
Key sources
- SOURCE@tencenthunyuan“Built for productivity”x.com
- SUPPORT@xeophon“putting it roughly on par with GLM-5.3 (!)”x.com
- SUPPORT@sgl_project“day-0 serving support from SGLang!”x.com
- SUPPORT@arena“this is an early AutoEval score, in which a Reward Model trained on Arena's human preference data casts automatic votes”x.com
- SUPPORT@kimmonismus“Hy4 coordinated several Codex sessions in parallel, evaluated their results and adjusted the research direction”x.com
- SUPPORT@testingcatalog“new open weight model under Apache License 2.0!”x.com
- SUPPORT@vllm_project“256 routed experts plus one shared”x.com
- SOURCEhuggingnewshuggingnews.com