Tencent releases Hy4 with first-time computer use, up 19.1% on GameDev bench
dl_weekly · x · 2026-11-15
Tencent officially released Hy4, adding computer-use capabilities for the first time, combining visual understanding, reasoning, and tool-calling for long-horizon agentic performance.
- The prior Hy4 Preview (Aug 28) had 770B parameters with 49B active; both generations offer a 1M-token context window, on par with open-weight frontier models from Mistral, Moonshot, and Z.AI
- Tencent says it substantially scaled pre- and post-training, with gains in instruction following, long-context understanding, agentic long-horizon tasks, and deep research, while fixing the previous version's excessive reasoning and over-verification during "thinking"
- Optimized for real-world productivity: web development, document summarization, visual creative tasks, presentations; front-end and game development improved, up 19.1% over Preview on GameDev bench
Weights and system cards are promised within days.
More from Models
- Google AI Mode now enforces usage limits, users surprised by rate-limit popup — elvano124 · 2026-11-22
- Sol 6.2 now refuses reverse engineering that Sol 6.1 happily did for a month — cardak98 · 2026-11-22
- Anthropic's record ~$81B IPO, GPT-6.2 pricing, Qwen4-Max and Copilot Autopilot: week in review — ksraj1001 · 2026-11-22
- Reddit rant: Frontier AI labs are training models to imitate lazy engineers — RossPeili · 2026-11-22
- MBZUAI open-sources PAN 2, a 29B-parameter general-purpose world model — LinusEkenstam · 2026-11-22
- Inkling Mini turns out surprisingly fluent in Russian, sparking data recipe questions — teortaxesTex · 2026-11-22