Tencent open-sources Hunyuan Hy4-preview: 770B params, 1M context, ~#3 among open models
机器之心 · wechat · 2026-08-28
Tencent Hunyuan released and open-sourced Hy4-preview: 770B total parameters (MoE, 49B active), 1M context window, weights and FP8 version under Apache 2.0. A big step from Hy3-preview (295B/21B/256K). API pricing: ¥6/M input tokens, ¥18/M output, ¥0.3/M cached input.
Benchmarks: Terminal-Bench 2.1 up from 70.8 to 85.4, DeepSWE 28.0 to 64.3, GPQA Diamond 92.3; it scores 1633 on CodeArena:WebDev (AutoEval), roughly #5 overall and #3 among open models. Positioned "for productivity," it targets long-horizon software engineering, multi-file office analysis, Unity/Unreal game prototyping from a single prompt, and scientific research; 163 internal experts blind-rated 203 engineering tasks at 2.99/4, slightly above GLM5.3 (2.92) and Kimi K3 (2.94).
Notably, the model participated in Hunyuan's own training pipeline, data strategy, and kernel optimization, lifting end-to-end inference throughput 31.8% over baseline.
Related event: Tencent Open-Sources Hunyuan Hy4 Preview: 770B MoE with 1M Context(16 posts)→
More from Infra
- Open-source Discord AI assistant Zauq: Multi-model routing & Docker sandbox — rar_file-exe · 2026-08-28
- Utility proposes rate cuts due to data center profits, saving users $100/year — bennash · 2026-08-28
- Why your local LLM feels dumber than hosted versions: Implementation details — JeremyCMorgan · 2026-08-28
- GTA VI Size Rumored to Range Between 0.748TB and 2TB — PtrPomorski · 2026-08-28
- Own a frontier AI model running locally in just 5 hours — MaziyarPanahi · 2026-08-28
- Optimal parameter settings for local deployment of Qwen3.8 Flash Next shared — Motor_Ad16 · 2026-08-28