Tencent Open-Sources Hunyuan Hy4 Preview: 770B MoE with 1M Context

Tencent released Hunyuan Hy4 preview on 08-28, first on WorkBuddy, then integrated into OpenCode Go, and open-sourced it under Apache 2.0 on 08-30 with vLLM support from day one, validated on NVIDIA GPUs. The model is a 770B-total / 49B-active MoE with 1M token context and a native MTP layer for speculative decoding, and is described as Tencent's strongest model to date.

Confirmed

• Official posts (@TencentHunyuan) highlight strong performance in research synthesis, Office workflows (cross-file analysis, Excel accuracy, PPT layout), front-end development, and game prototyping, optimized for Coding Agent scenarios.
• Internal blind test: 163 experts over 203 real engineering tasks scored Hy4 preview 2.99/4, beating GLM-5.3 (2.92) and Kimi K series.
• @AlchainHust reported an Arena coding score of 1633, on par with Grok-4.6 and Claude Fable 5; official evals claim metrics approaching GPT 5.6 Sol.
• @vista8's 8-project test found major front-end aesthetic gains, stable long-horizon tasks, and strong office usability, including audio-driven animation effects.
• An invited tester (retweeted by @TencentHunyuan) judged overall capability close to GLM-5.3 and above the DeepSeek V4 series, with clear coding improvements.
• A developer built a fully playable browser game in WorkBuddy using the model.
• Pricing starts low for input tokens (original post truncated at "$0…").

Why it matters

• Apache 2.0 open-sourcing of a 770B flagship MoE with 1M context and MTP speculative decoding gives developers a self-hostable top-tier Chinese model for long workflows and agents, while low pricing may intensify domestic model price competition.

2026-08-28 ~ 2026-08-30 · 12 related posts

• [source] Hands-on with Tencent Hy4 preview across 8 projects: better frontend taste, stable long-horizon tasks — vista8 · 2026-08-28
• Tencent's Hunyuan Hy4 Preview Launches: Targeting Kimi and GLM for Productivity Tasks — AlchainHust · 2026-08-28
• [source] Tencent Hunyuan Hy4 Preview Shows Strong End-to-End Agent Performance — TencentHunyuan · 2026-08-28
• Tencent's Hunyuan Hy4 Preview excels in complex coding and planning workflows — Aiden_Tech_Ai · 2026-08-28
• Tencent Hunyuan Hy4 Preview Integrates OpenCode Go with 1M Context — TencentHunyuan · 2026-08-29
• Hands-on Tencent Hy4 Preview: Matches GLM-5.3, Major Coding Gains — TencentHunyuan · 2026-08-29
• Tencent Releases Hy4 Preview: 770B Parameter MoE Model for Coding and Research — pstAsiatech · 2026-08-30
• Tencent Hunyuan Hy4-preview runs in vLLM day 0: 770B MoE with 1M context — TencentHunyuan · 2026-08-30
• [source] Tencent open-sources Hunyuan Hy4: 770B MoE with 1M context — TencentHunyuan · 2026-08-30
• Tencent releases Hy4 preview: 49B active params, 1M context, open source — Div_pradeep · 2026-08-30
• Tencent Hunyuan Hy4 Open Sourced: 770B Params, Beats Kimi K3 in Internal Tests — nikola_mr64990 · 2026-08-30
• Tencent open-sources Hy4 preview with 770B params and 1M context — testingcatalog · 2026-08-30