Tencent Hunyuan Hy4 official release: first native vision, OSWorld 2.1 at 78.9

AGI Hunt · wechat · 2026-11-04

Tencent has released the official version of Hunyuan Hy4, claiming first-tier open-source status, with weights and a system card to be open-sourced soon. The headline change: native visual understanding plus computer use (CUA) — the model can see the screen and operate mouse/keyboard alongside CLI and tool calls.

Key numbers: Terminal-Bench 4.0 up from 26.3 to 44.4, 3DArena-Modeling from 23.4 to 37.1, and OSWorld 2.1 at 78.9, beating GPT-6.1 Sol's 75.2. On third-party AgentArena, +4.7% net gain across 8,500+ real long-horizon agent tasks (rank 25 → 14) at $0.55 per task, entering the Pareto frontier. In finance, a blind test of 579 real tasks by 213 experts (most from top institutions) ranked it first among open models. It also helped Peking University mathematicians find a counterexample to the Méndez–Pinto–Villarreal conjecture, with Lean-verified proofs. Pricing is unchanged; the author's end-to-end interactive picture-book iOS app built with WorkBuddy+Hy4 felt "not inferior to Codex", though thinking is slow and top closed models still lead on aesthetics and very long tasks.

Original post →

Tencent Hunyuan Hy4 official release: first native vision, OSWorld 2.1 at 78.9

More from coding & agent