Tencent open-sources Hy4 preview, a 770B-parameter MoE model for coding and research

Tencent released an open-source preview of a new large language model called Hy4 preview on August 28, 2026, publishing the weights on Hugging Face, ModelScope, GitCode, and CNB. The model is aimed at software engineering, office analysis, game development, and scientific research tasks rather than general chat.

What's new

According to the model card, "the model comprises 770B total parameters, of which 49B are activated per token. The backbone consists of 78 layers, where the first layer uses a standard dense FFN and the remaining 77 layers replace it with MoE, each containing 256 routed experts and 1 shared expert." That sparse mixture-of-experts design lets Tencent ship a very large model while keeping per-request compute closer to a much smaller dense model.

Tencent built the model by pulling in domain specialists from inside the company — software engineers, game developers, finance analysts, and security staff — and training on data drawn from their actual work. The stated goal is a model that gets "meaningfully further on tasks like software engineering with better understanding, planning, debugging, and verifying of long-horizon development tasks," rather than one tuned mainly for short conversational exchanges.

Tencent also shipped a lower-precision Hy4-preview-FP8 variant alongside the full-weight release, giving developers a cheaper option to run inference on. The company acknowledges this is an early release: the model can sometimes take longer than necessary to work through complex questions and may over-verify its own answers, and the team notes "there is real headroom left in both pre-training and post-training" for future versions.

Context

Hy4 preview extends Tencent's Hunyuan model family into a new, more agentic-coding-focused generation, arriving alongside a wave of open-weight releases from Chinese labs this year — DeepSeek, Alibaba's Qwen, Moonshot's Kimi, and Zhipu/Z.ai's GLM have all shipped large open models aimed at competing with closed frontier systems on coding and agentic benchmarks. Tencent says it plans to integrate Hy4 into its own products, including its CodeBuddy coding assistant and WorkBuddy office-productivity tool, mirroring how other labs pair an open-weight release with an in-house commercial product built on top of it.

Why it matters

A 770B-parameter open-weight model with only 49B active parameters per token is a significant release even in preview form — it puts frontier-scale MoE architecture directly in developers' hands to fine-tune or self-host, at a fraction of the inference cost of a fully dense model that size. The specific focus on long-horizon software engineering tasks, rather than general assistant capability, signals where Tencent sees the most competitive pressure and commercial upside: agentic coding tools are one of the most contested categories in AI right now, and an open, self-hostable option from a major Chinese tech company adds real competition for teams that can't or won't route code through a closed API.