Tencent launched Hy4 preview, its open-weight large language model (LLM) with a total of 770 billion parameters and 49 billion active parameters, placing it further into the busy competition among Chinese AI laboratories. Unveiled on August 28 2026 its mixture-of-experts architecture boasts a context window of over a million tokens. Tencent claims it as its strongest model to date, with a difference between the previous Hunyuan generations as the largest has measured so far. The architecture aims to achieve the effects of scale, while not requiring the full inference cost of a dense 770-billion-parameter network: only some fraction of the experts fire for each token, so compute per request remains comparable to that of a medium-sized dense model.
Tencent’s implementation includes 78 layers, with the first one realized as a typical dense feed-forward network, and the remaining 77 realized using MoE routing through hundreds of experts and one shared expert. The company puts out the model alongside a multi-token prediction layer which is said to speed up decoding. Weights are available via Apache 2.
0 license on Hugging Face and associated places, though the pretraining data and complete training recipe are not. Tencent’s target for the model is for long-horizon work, not word-mongering chat conversations. It passes of software engineering, including combining planning, debugging and verification during long hours; transforming disorganized files into reports, spreadsheets and slide decks through heavy documentation and office work; diagramming prototypes from natural-language prompts; and (HT) scientific investigation. In an internal blind review, 163 Tencent experts scored 203 engineering tasks: the Hy4 preview scored an average of 2.99 out of 4. edging out GLM-5.3 at 2.
92 and Kimi K3 at 2.94. Those are company-run than independent third-party leaderboards, not exactly a secret that the company is not hiding.] The launch comes as Tencent speeds up its AI cadence following the hiring of ex-OpenAI researcher Yao Shunyu and the release of Hunyuan 3.0 earlier this year. Hy4 preview is already integrated into Tencent services such as CodeBuddy WorkBuddy Yuanbao and ima, with API access via Tencent Cloud TokenHub and OpenRouter. The company launched with two weeks free trial on WorkBuddy and CodeBuddy, and published fairly aggressive API prices, listing cached input far below uncached input plus output.
Tencent also points out that the preview is still a work in progress: it can overthink on complex questions and over-verify its own answers. The more distant context is an open-weight Chinese outlier WAVE which in 2026 has again leapfrogged U.S. labs on raw number of parameters. An open license for a 770B flagship is both a technical and a commercial message: it’s large enough to allow developers to fine-tune and deploy it themselves, while maintaining the secret sauce inside the Tencent productivity chrome. Though if Hy4 preview turns into a stalwart open-source over time will depend less on the headline number of parameters, but on how the active-parameters design, how a million-token context, and how real-task fine-tuning stand the test of independent users running its codebase through coding, research, and office workloads outside the Tencent bubble.