A model that can read a chart, click through a design app, and finish a multi-step task on its own — that is the new bar for a large model (大模型) in late 2026. On 4 November, Tencent released Hunyuan Hy4, its newest flagship language model, and for the first time placed the Hunyuan family among what the company calls the strongest open-source models.
Key takeaways
- Hy4 scales the Hunyuan series to 770 billion parameters; the prior Hy3 series sat under 300 billion.
- It adds visual understanding for the first time versus Hy4-preview, and improves instruction following, long-context, and long-horizon agentic (智能体) ability.
- Hy4 supports ComputerUse — operating a PC through command line and GUI — with specialized training on Blender and CAD.
- Tencent plans to open-source the weights and publish a model system card.
- Since rebuilding its infrastructure in February, Hunyuan has shipped a major version roughly every two months.
What changed versus the preview
Compared with Hy4-preview, Hy4 significantly expanded both pretraining and post-training data and scale. The notable first is visual understanding: the model can now handle fine-grained image processing, complex chart analysis, and math visual reasoning, not just text. Long-context and long-horizon agentic capabilities — the ability to plan, call tools, and run a cross-application task to completion — also improved materially.
Parameter scale is the blunt headline. The Hy3 series used fewer than 300 billion parameters; Hy4 reaches 770 billion. Tencent frames the climb as part of a steady cadence: after rebuilding the training stack in February, the team ships a big version about every two months, opening a Preview first and then a tuned official release based on user feedback. Hy4-flash, a lower-cost variant, is expected soon.
Why ComputerUse matters
The headline capability for real work is ComputerUse. Hy4 can operate a computer by combining command-line and graphical-interface actions, handling everyday and professional tasks. Tencent says it ran targeted training for Blender, CAD, and other professional software, plus long-horizon cross-app workflows, lifting the model's planning and execution on specialized jobs. Tied to WorkBuddy's enterprise WeChat suite, Hy4 can connect chat history, email, documents, calendars, and meetings — and when several projects run in parallel, it can sort out each one's progress, flag tasks to follow up, and remind the user.
Tencent describes the direction as a shift from a tool toward a teammate: agents (智能体) that do not just generate text but complete tasks inside the user's actual software.
The open-source move
Tencent says Hy4 will open-source its weights and publish a model system card, with an explicit focus on responsible release — flagging network and biosafety risks. That puts Hunyuan in direct competition with the open-weight arms race led by Qwen, DeepSeek, and others, where the contest is no longer only model quality but the size and loyalty of the developer ecosystem built on top.
Hunyuan continues its stated low-cost, inclusive路线 (route), offered for global developers to call. Whether "strongest open-source" holds up is something the community will test once weights are out; Tencent's own positioning is a claim, not yet a third-party verdict.
The ecosystem race is the real stakes. Open weights do not earn revenue directly, but they pull developers, fine-tuning jobs, and inference traffic toward the issuer's cloud and tooling — the same logic that has made Qwen and DeepSeek household names among builders. For Tencent, a company historically stronger in consumer products than in developer infrastructure, Hy4's open release is also a bid to plant a flag in that builder community. The cost of that strategy is real: open-sourcing narrows the gap between a flagship model and free alternatives, so the moat shifts to serving, safety tooling, and enterprise integration rather than the weights themselves.
What readers can do now
- Wait for the actual weight release and license before planning production use. "Will open-source" is a commitment, not a download yet — read the terms when they land.
- If you work in design, CAD, or office automation, test Hy4's ComputerUse on real multi-step jobs once access opens, rather than judging from demos.
- Treat "strongest open-source" as a positioning statement, and benchmark against Qwen and DeepSeek on your own workloads.
Honest limitations
This article draws on Tencent's official site, People's Daily (人民网), and CCTV (央视网) reporting. The 770B parameter figure, the Hy3 sub-300B comparison, the first-time visual understanding, ComputerUse, and the open-sourcing plan are reported per those sources. Hunyuan's absolute ranking among open models comes from Tencent's own framing and media paraphrase, not an independent benchmark re-run here. The exact release date and license terms for the open weights were not finalized at writing, and I have not covered Hy4-flash specifics or Hunyuan's consumer-app integrations beyond WorkBuddy.