Tencent open-sources Hy4 Preview, a 770B MoE model it says helped build itself

Tencent has released open weights for Hy4 Preview, a mixture-of-experts LLM with 770B total parameters (49B active per token) and a context window exceeding 1M tokens, positioned squarely at productivity work: coding, office and document tasks, and scientific research. Beyond the open-source download, it’s wired into Tencent’s own apps — WorkBuddy, CodeBuddy, Yuanbao, and ima — and reachable via API through Tencent Cloud TokenHub and OpenRouter. The commercial hooks are aggressive: two weeks free on WorkBuddy and CodeBuddy, extended free Hy3 access through September 30, and low API pricing at roughly $0.83 per million input tokens and $2.50 per million output tokens.

Tencent’s performance claims are notable but self-reported. In an internal blind evaluation of 203 engineering tasks graded by 163 experts, Hy4 Preview averaged 2.99 out of 4.00, edging out GLM-5.3 (2.92) and Kimi K3 (2.94) — a thin margin from a benchmark the vendor designed and ran, so it warrants independent verification before being taken as a ranking. The pitch leans heavily on real-world scenarios co-designed with Tencent’s own products, including front-end development, financial and data analysis, cross-document workflows, and generating playable game prototypes from a single prompt.

The most striking assertion is about the model’s role in its own development. Tencent says Hy4 Preview participated in automated optimization of its training methods, data strategies, evaluation frameworks, and low-level operators — proposing approaches, running experiments, and iterating on the results — which it frames as an early recursive self-improvement loop. It also claims the model autonomously profiled and tuned its inference stack (operator fusion, communication optimization) for a 31.8% end-to-end throughput gain. These are the claims most worth scrutiny: self-improving AI is a heavily hyped frontier, and Tencent offers the numbers without external validation.