OPEN LINK ↗
// 1h agoMODEL RELEASE
Tencent Hy4 Preview Goes Open Source
Tencent’s new Mixture-of-Experts model packs 770B total parameters, activates 49B per token, and supports a 1M-token context window. Released under Apache 2.0 with FP8 support, it targets coding, office work, game development, and scientific research.
// ANALYSIS
Hy4 preview is a serious open-model release, but its capability claims remain preliminary until independent evaluations arrive.
- –MoE routing keeps per-token computation at 49B active parameters, while the 770B capacity raises substantial memory and serving requirements.
- –Native 1M context could enable repo-scale coding, cross-document analysis, and long-running agent workflows.
- –Apache 2.0 weights, FP8 quantization, vLLM, and SGLang support make experimentation practical for well-equipped teams. [Hugging Face model card](https://huggingface.co/tencent/Hy4-preview-FP8)
- –Tencent reports a 2.99/4 internal blind-test score versus 2.92 for GLM-5.3 and 2.94 for Kimi K3, but those results are not independent public benchmarks. [Test details](https://www.chooseai.net/news/6281/)
// TAGS
hy4-previewllmopen-sourceopen-weightsmoelong-contextai-coding
DISCOVERED
1h ago
2026-08-28
PUBLISHED
2h ago
2026-08-28
RELEVANCE
10/ 10
AUTHOR
XQOPTRX