Tencent HY4 Tested: 770B MoE Open-Weight Model Challenges Claude and GPT
WorldofAI · youtube · 2026-09-01
Tencent released HY4 Preview, an open-weight model: 770B-parameter MoE with 49B active parameters and a 1M-token context window.
YouTuber WorldofAI ran real-world tests across coding, 3D game development, and long-horizon agentic tasks, comparing it against Qwen 3.8, GLM 5.3, DeepSeek V4 Flash Vision, Kimi K3, Claude Opus 5, and GPT-5.6. HY4 produced a 3D Spider-Man-style game, zombie FPS environments, a Unity penguin game, and an ISS simulation.
The community also compressed the model from 1.5TB to 200GB with surprisingly little benchmark degradation. The review concludes HY4 may be the best open-weight model right now, fully free.
More from Models
- Yoav Goldberg questions the confusing name 'helpful-only models' — yoavgo · 2026-09-01
- Community Uncensored Fine-tune of Qwen3.8-27B Trends on Hugging Face — DavidAU · 2026-09-01
- ExLlamaV3 update: MoE expert CPU offload, GLM-5.3-Flash, self-calibrated quants — Unstable_Llama · 2026-09-01
- Qwen3.8 pMLX Engine: Runs on 12GB RAM at 10 tok/s — EyalToledano · 2026-09-01
- Dev Reports Sol Built a Useful Model Probe Tool and Handled Refactors Well — voooooogel · 2026-09-01
- Dev says GPT-5.6 Sol refuses to remove code, echoing Redwood's "AIs are misaligned" post — osmarks1 · 2026-09-01