Tencent open-sources Hy4: 770B-parameter native multimodal MoE with 1M context
ResearchCrafty1804 · reddit · 2026-11-20
Tencent released the weights of Hy4, its natively multimodal flagship MoE model with 770B total / 49B active parameters and 1M-token context.
- Available on GitHub and Hugging Face, with GGUF quantized versions and a system card also published.
- Official benchmarks pit Hy4 against Claude Opus 5.5, GPT 6.1 Sol, Kimi K3.1 Pro, GLM 5.3, and DeepSeek-V4.1-Flash across coding and agent evals (Terminal Bench 4.0: 66.8, OSWorld 2.1: 78.9), where it leads or ties on several boards, some evaluated for the first time at release.
Related event: Tencent Open-Sources 770B Multimodal Flagship Hy4(4 posts)→
More from coding & agent
- Procedural Gojira Built with Gemini and Three.js TSL — majidmanzarpour · 2026-11-20
- User finds their AI agent quietly ate 54GB of disk with unused git worktrees — itsOmSarraf_ · 2026-11-20
- Harness engineering gap: Provider harness 39.0% vs Mini-SWE 12.8% on same model — andrew_n_carr · 2026-11-20
- OpenAI hiring for computer-use agents team as Apple lists similar ambition — yungcontent · 2026-11-20
- Codex subscription works with your own cloud harness, and one dev wants OpenAI to win for it — gonzohst1 · 2026-11-20
- Bug Hunt chart: score vs subscription cost, Gemini 4 Neon xhigh tops at 63 for $1.80 — PawelHuryn · 2026-11-20