Tencent released Hunyuan Hy4 in general availability on November 4, using language bolder than any previous launch: the Hy series, it says, has joined the ranks of the strongest open models.

The parameter count backs part of that. Hy4 reaches 770B, up from under 300B in the Hy3 series, with 49B active parameters and a context window exceeding 1M tokens. A next-generation Hy5 is expected to scale further.

Two firsts

The first is vision. Hy4 is the series flagship that can see: visual understanding, reasoning and tool calling combine for fine-grained image handling, complex chart analysis and mathematical visual reasoning.

The second is computer use. The model can operate through both command line and graphical interfaces, with targeted training for professional software such as Blender and CAD, and for long-horizon cross-application tasks.

Read together, the two capabilities point in one direction: from a model that talks to a model that acts.

How to read the benchmarks

Three published comparisons are worth attention.

In a crowdsourced test of 579 real financial tasks, Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68.

On the IRAB internal benchmark from financial agent company Xuntu Technology, Hy4 improved 12.03 percentage points over the preview version.

In office scenarios, average quality scored 83.18 with a 96.15 percent task pass rate, ranking first domestically in 9 of 12 scenarios.

Tencent also concedes the gap: despite being among the strongest open models, Hy4 still trails top closed systems substantially. It beats several domestic models on coding, automation and computer-operation benchmarks, but remains behind GPT-6.1 Sol and Claude Opus 5.5.

That kind of self-reporting makes the numbers more credible, not less.

The cadence changed

Tencent Cloud says Hunyuan has shipped a major version roughly every two months since rebuilding its infrastructure in February 2026, following a fixed pattern of preview first, GA later, feeding real-world usage back into development.

The Hy4 preview in late August was a product of that cadence, positioned then as "top tier among open models." Two months later the claim has moved to "among the strongest open models."

Where to use it

Hy4 is live in Tencent Cloud Tokenhub, WorkBuddy/CodeBuddy, Yuanbao, the WeCom assistant and ima. The API is available through OpenRouter, Hermes and OpenCode. Tencent properties including QQ, Honor of Kings, Peacekeeper Elite, WeRead and Tencent News are integrating it.

Tencent says it will release model weights and a system card soon, framed as "more responsible open source."

What it means for developers

If the weights do ship, 770B is not something most teams can run. The realistic paths are API calls or distillation into smaller task-specific models.

The two capabilities worth evaluating are narrower:

Can vision reliably handle Chinese charts and screenshots? That is the highest-frequency office need in this market.

Does computer use survive long-horizon tasks in real software? The gap between demo and production usually appears here.

Tencent's move here is tying model capability to product distribution. For users, parameters and leaderboards are references. Three things decide adoption: pass rate on your tasks, cost per task, and whether the tooling ecosystem cooperates. All three require your own testing.