Illustrative image: servers in a data centre, where models like Hy4 run. Hy4 is software, not a device, so there is no product photograph for this release. Image: Wikimedia Commons / BalticServers, CC BY-SA 3.0.
TL;DR — Tencent Hunyuan Hy4 AI model explained in one line: on 4 November 2026, Tencent released Hy4, a 770-billion-parameter model that can see images and operate a computer — but the open weights are not out yet, so today you can only use it as a hosted service, not download it.
📌 KEY HIGHLIGHTS
- Released today: Tencent's Hunyuan team released the official version of Hy4 on Wednesday, 4 November 2026, succeeding the Hy4 preview open-sourced on 28 August 2026.
- Big model: 770 billion total parameters in a mixture-of-experts design, with 49 billion active parameters and a context window beyond 1 million tokens, per the model's listings.
- Two firsts for Hunyuan: visual understanding (fine-grained image processing, chart analysis, visual maths) and Computer Use — operating a machine through its graphical interface and command line.
- Available now, as a service: Tencent Cloud TokenHub, Yuanbao, ima, CodeBuddy/WorkBuddy, WeCom's Dayuan assistant, and APIs via OpenRouter and others.
- The honest catch: the model weights are not downloadable yet — Tencent says they will be open-sourced soon, and a lighter Hy4 Flash is also coming, with no date for either.
Tencent Hunyuan Hy4 AI model explained: what happened?
Tencent announced Hy4 and its companion Hy4 Flash at its Global Digital Ecosystem Summit on 29 October 2026, and the official Hy4 version followed on 4 November. It is the next step in a fast-moving open-model race: the Hy4 preview from late August was already a 770-billion-parameter open-weight model, and the official version expands both pre-training and post-training scale on top of it, with sharper instruction following, longer-text understanding, longer-horizon agent work and deeper research, according to Tencent.
The context matters. In the same week, Chinese labs have been shipping at a striking pace — Meituan's LongCat-2.5, SenseTime's SenseNova 6.8 Flash, and Zhipu's GLM-5.3 all landed within days — and the open-weight frontier is now crowded with genuinely strong models. Tencent's pitch is that Hy4 has, for the first time, joined the ranks of the most powerful open-source models. That is Tencent's framing, and it rests on Tencent's own evaluations: no independent benchmark results for the official Hy4 version exist yet, and this post does not present vendor tests as independent proof.
What is actually new: vision and computer use
Two capabilities are new to the Hunyuan line, and both are worth understanding separately — because together they move Hy4 from a text assistant toward something that can work from what it sees and act on a machine.
| Capability | What it means | Status |
|---|---|---|
| Visual understanding | Reads images in fine detail, analyses complex charts, reasons about mathematical visuals — combining perception, reasoning and tool use | Live in Hy4 |
| Computer Use | Operates a computer via graphical interface and command line, including Blender and CAD workflows and multi-app tasks | Live in Hy4 |
| Long-horizon agents | Multi-step work across applications — office collaboration, financial analysis, software development, cross-app automation | Live in Hy4 |
| Front-end and game development | Targeted optimisation for complex interfaces and engines (UE, Unity, Godot); Tencent reports GameDev +19.1% and GameLogic +9.8% over the preview | Tencent's own evaluation |
| Open weights | Downloadable, runnable, fine-tunable model — the part that makes an "open" model actually open | Promised soon, not out yet |
| Hy4 Flash | Lighter, speed- and cost-focused version inheriting Hy4's visual and agent capabilities | Coming soon, no date |
Tencent's own numbers give a flavour of where it thinks Hy4 wins. In a financial crowd-test run with Tencent Financial Technology, 213 experts (153 from leading outside financial institutions) scored Hy4 at 3.04 on 579 real financial tasks, ahead of Kimi K3.1 Pro (2.68), GLM-5.3 (2.68) and DeepSeek-V4.1-Flash (2.97) — first among open-source models in that test. In internal search testing, the hallucination rate fell from 15.76% in the preview to 5.45%, and guessing-driven errors from 23.1% to 5.6%. And in Tencent's TenureTrail scientific-research evaluation, Hy4 ranked first among open-source models. All of those are Tencent's figures, from Tencent-run tests — useful signal, not independent verification.
Illustrative image: code on screen. Hy4 is pitched at developers and agent workflows — coding, front-end work and multi-app automation — so a code screen is the closest honest visual. Image: Wikimedia Commons.
Why it matters for World Tech readers
The open-weight race is no longer about one lab. A developer choosing a model today can pick between Tencent's Hy4, Moonshot's Kimi K3.1 Pro, Zhipu's GLM-5.3 and DeepSeek's models — all open or open-promised, all with million-token context, all pushing into agentic work. That competition is good for everyone building on AI, because it pushes capability up and prices down: Hy4's API pricing stays at the preview's level (RMB 0.3 per million tokens for cached input, RMB 6 for regular input, RMB 18 for output, per reporting), which undercuts most closed frontier models by a wide margin.
For Indian developers, the practical angle is immediate: Hy4 is reachable today through APIs on OpenRouter and other gateways, and through Tencent's own products, without waiting for the weights. The same week's Indian AI story — see our IndiaAI EKA-1 DhÄ«r models India explainer — shows the other side of the same coin: sovereign models built for Indian languages and offline use, while Hy4 shows what the global open frontier now offers off the shelf. And the caution from our Next.js November 2026 security release explainer applies here too: a model that can use a computer is a model whose actions need watching — Computer Use is powerful precisely because it crosses from answering to doing.
One more thing Tencent is signalling: Hy4 is being prepared for Recursive Self-Improvement work — AI that helps build better AI — with its Hyra2.0 scaffolding, which Tencent says already trained a smaller "Jev" model that beats the official version on several evaluations, and produced two open-sourced science models, HyCell and HyPDE. Treat that as a research direction, not a capability claim: the interesting question is what independent benchmarks say once the weights actually land.
Should you care?
✅ Yes, if you build with AI models. Hy4's vision plus Computer Use is a genuinely new combination in the open-model race, and it is usable via API today. ⏳ Wait for the weights before calling it open. Until Tencent publishes them, nobody outside Tencent can download, run, fine-tune or inspect Hy4 itself — the "open-source" label is a promise, not yet a download.
FAQs
What is Tencent Hunyuan Hy4? Hy4 is Tencent's new-generation large language model, released on 4 November 2026 as the successor to the Hy4 preview from August 2026. It is a 770-billion-parameter mixture-of-experts model with visual understanding and Computer Use — the ability to see images and operate a computer — aimed at office, finance, coding and research work.
Is Hy4 open source? Not yet, in the downloadable sense. The Hy4 preview's weights were open, and Tencent says the official Hy4 weights and a system card will be open-sourced soon, but as of today you cannot download Hy4 itself. What you can use now is the hosted model through Tencent's products and APIs — so call it open-promised, not open-delivered, until the weights appear.
What can Hy4 do that Hy4 preview could not? Two things: see and act. The preview was text-focused; Hy4 adds visual understanding (detailed image reading, chart analysis, visual maths) and Computer Use (driving graphical interfaces and command lines, including Blender and CAD workflows), plus stronger long-horizon agent performance and lower hallucination in Tencent's internal tests.
How can I try Hy4 today? Through Tencent Cloud TokenHub, the Yuanbao and ima apps, CodeBuddy/WorkBuddy, the WeCom Dayuan assistant, or APIs on OpenRouter and other gateways, at the preview's pricing. If you need to run a model on your own hardware, wait for the weights — or use the Hy4 preview, whose weights are already open.
TAKEAWAY: Hy4 is Tencent's strongest open-model bid yet — vision, computer use and 770 billion parameters, live today as a service. The verdict that matters comes when the weights land and independent benchmarks run: until then, enjoy the hosted model, but hold the "open" label lightly.
Follow World Tech on TechByCountry for tomorrow's 6 PM update.
📸 Get tech in the last 24 hours on Instagram — follow @_techinlast24hrs for daily posts (9 AM–9 PM) and reels at 12:30 PM & 8:30 PM.
Sources
- Zhitong Finance via Longbridge — Tencent Hunyuan Officially Releases Hy4 (4 Nov 2026)
- ChainCatcher — Tencent Hunyuan Releases Hy4 (4 Nov 2026)
- dev.to — Tencent's Hy4 Can See and Use Your Computer, Explained Honestly (4 Nov 2026)
- wedoany — Tencent Releases Official Version of Hunyuan Hy4 (4 Nov 2026)
- HiTechExpert — Tencent Unveils Hy4 (4 Nov 2026)
Comments
Post a Comment