Tencent open-sources its Hy4 preview model

Tencent released and open-sourced Hy4 preview, the latest model from its Hunyuan team, in a company post dated August 28. Read the headline number carefully. The 770 billion parameters are the model's total, not a dense count. Hy4 preview is a Mixture-of-Experts model that activates about 49 billion parameters per token, with a context window the company puts at more than 1 million tokens.

Open weights under Apache 2.0

This is a weights release under a permissive license, not a closed API-only launch and not a paper with nothing to download. Tencent published the weights under the Apache License 2.0 and posted them to Hugging Face, ModelScope, GitCode and CNB. The GitHub model card describes 78 layers, 256 routed experts plus one shared expert, and the top-8 routed experts activated per token. That sparse routing is why a 770B-total model can run while touching only 49B parameters on any given pass, and it is the detail most one-line takes drop when they call this a flat "770B model."

Availability is broad from day one. Beyond the open weights, Hy4 preview is wired into Tencent's own WorkBuddy, CodeBuddy, Yuanbao and ima apps, with API access through Tencent Cloud TokenHub and OpenRouter. Tencent lists TokenHub prices at $0.834 per million input tokens, $2.501 per million output, and $0.042 per million cache-hit tokens. The company is also running a promo: Hy4 preview is free on WorkBuddy and CodeBuddy for two weeks from launch, and free access to the older Hy3 on both apps runs until September 30.

Read the benchmark as Tencent's own

Tencent quotes an internal blind evaluation, 163 experts scoring 203 engineering tasks, in which Hy4 preview scored 2.99 out of 4.00 against GLM-5.3 at 2.92 and Kimi K3 at 2.94. Treat that as a vendor number, not an independent bench. It is Tencent grading a Tencent model, the margins are narrow, and none of it converts into "beats GPT-5." The company frames the release in its own words, saying Hy4 preview "was expanded significantly in model size, context length, and data volume," and reports a 31.8% gain in end-to-end inference throughput over its baseline from its own serving work. A further batch in the Hy4 series is described as expected soon. TechNode carried the same figures.

The takeaway

If you are deciding whether to test this, do not price it as a 770B dense model you cannot run. Price it as a 49B-active MoE you can pull under Apache 2.0 and serve yourself, or rent on TokenHub at well under a dollar per million input tokens. Download the weights or take the two free weeks on CodeBuddy, run your own coding evals rather than trusting the 2.99-against-2.92 internal score, and judge from that. The open license is the part worth acting on here, not the leaderboard line.

For related context, see our coverage of OpenAI winding down its Cursor model contract, Nvidia pausing its AI cloud revenue-share deals, and Anthropic's model hardware standard.