Tencent open-sources its Hy4-preview language model

Tencent released Hy4-preview under an Apache 2.0 licence, a mixture-of-experts model with 770 billion total parameters, 49 billion active per token, and a context window beyond one million tokens, succeeding Hy3, the 295-billion-parameter open model the company had released in July.

Tencent reported an internal blind evaluation in which 163 outside experts rated the model against rivals on 203 real-world engineering tasks, putting Hy4-preview marginally ahead of both GLM-5.3 and Kimi K3 — average scores of 2.99 against 2.92 and 2.94 out of 4 respectively, a margin the company itself did not present as decisive. On Terminal-Bench 2.1, a benchmark of agentic command-line tasks, Tencent reported a score of 85.4. The company disclosed two limitations directly: a tendency to spend longer than necessary reasoning through complex tasks, and to over-verify its own completed work, adding that it saw “real headroom left in both pre-training and post-training.”

The model is free to use for two weeks through Tencent’s WorkBuddy and CodeBuddy tools and the Yuanbao and ima consumer apps, after which list pricing is set at $0.834 per million input tokens and $2.501 per million output tokens via Tencent Cloud and OpenRouter. The release continued a run of large open-weight Chinese models — Hy3, GLM-5.3 and Kimi K3 among them — increasingly benchmarked directly, and favourably, against each other and against closed alternatives by their own developers.