Tencent releases Hunyuan Hy4-preview, a 770B open-weights MoE claimed to lead SWE-bench Pro and jump on WebDev arena
2026-08-28
Per Latent Space's AINews roundup, Tencent Hunyuan (@TencentHunyuan) released Hy4-preview, a 770B-total/49B-active-parameter open-weights mixture-of-experts model with a 1M-token context window, explicitly framed by Tencent as 'open source frontier.' Aggregated social reports cited by AINews say the model placed around #5 on a 'Code Arena: WebDev' leaderboard via AutoEval — a +115 point jump over predecessor Hy3 — and that Cline's team said it leads on a benchmark called 'SWE-bench Pro.' Tencent also claimed Hy4 can coordinate multiple Codex sessions in parallel for research workflows. vLLM's team described its serving design as 256 routed experts plus 1 shared, with only 21 of 78 layers computing their own sparse index while others reuse it, plus an embedded 10B-parameter MTP layer with draft depth 3.
Significance 3: A new large open-weights MoE from a major Chinese tech company would be a significant new entrant if corroborated, but every claim here is social-media-relayed via a single curator digest with no primary Tencent source or independent benchmark reproduction, so significance is held down per the facts-vs-opinion rule.
Implications · machine-drafted, not owner judgment
If corroborated, Tencent joins DeepSeek, Qwen, and Z.ai as a fourth credible Chinese open-weights frontier contender within the same month, which would further narrow the open/closed gap the open-weights-one-generation-behind thesis tracks and give enterprises even more sovereign/self-hostable options at the high end. The claimed SWE-bench Pro lead and multi-Codex-session coordination are exactly the kind of agentic-coding claims that matter for the coding-agent verdict, but none of it is independently verified yet.
- A primary Tencent Hunyuan announcement or model card
- Independent SWE-bench Verified or Terminal-Bench scores for Hy4-preview
- Weights actually published (this is a 'preview'; confirm full release)
Sources
- curator [AINews] OpenAI shuts off Cursor retrieved 2026-08-29