Tencent Drops Hy4 — 770B MoE, 1M Context, Open Source

Yesterday, Tencent open-sourced Hy4 preview, their next-generation Hunyuan model — and the numbers are worth paying attention to.

770B total parameters, 49B active (MoE architecture). 1M+ token context window. Released under an open-source license and available globally. API pricing at $0.834/M input tokens and $2.501/M output tokens.

This isn't another press release model. The blind eval results Tencent cites — across coding, office productivity, game development, and scientific research — suggest Hy4 lands in the same tier as the best open-weight models available today. Notably, it powered a playable game prototype from a single natural-language prompt and showed strong performance on long-context development tasks.

But the most interesting detail is buried deeper: Hy4 contributed to its own development. Tencent reports the model was used to automatically optimize training methods, data strategies, evaluation frameworks, and inference bottlenecks — including autonomous analysis of operator fusion and communication optimization in its own serving stack.

This is the real signal. We're past the point where models are merely the output of human R&D pipelines. Hy4 participated in designing itself.

The context story also matters. A 1M+ token context window in an open model at this scale makes long-document analysis, full-repo coding, and multi-hour conversation histories genuinely practical — not just a benchmark checkbox.

Tencent is offering free access via WorkBuddy and CodeBuddy for two weeks, and the open weights are available now. Between this, DeepSeek's sustained lead on reasoning, and the Llama 4 ecosystem maturing, the open-weight frontier is more crowded — and more capable — than it's ever been.