Tencent has released Hy4, its new-generation large language model — and this one is different from its predecessors in two ways that actually matter to developers. It is the first model in Tencent's Hunyuan line that can see: fine-grained image processing, complex chart analysis, mathematical visual reasoning. And it is the first that can use a computer: operating a machine through its graphical interface and command line, combining perception, deep reasoning and tool invocation.
Hy4 succeeds the Hy4 preview from late August 2026, and the scale jump is large: 770 billion total parameters in a mixture-of-experts architecture, up from Hy3's 295 billion — a 2.6x increase. It is available now through Tencent's own surfaces — Tencent Cloud TokenHub, WorkBuddy/CodeBuddy, Yuanbao, ima, and the WeCom intelligent assistant Dayuan — and through APIs on OpenRouter, Hermes and OpenCode. The weights themselves are not out yet: Tencent says they will be open-sourced soon, and a lighter Hy4 Flash version focused on speed and cost is also coming soon. Neither has a date in the reporting, and this article does not invent one.
That distinction — available now as a service, not yet open as weights — is the first thing to keep straight, because much of the coverage of Hy4 leans on Tencent's framing of it as an open-weight model. Right now, what you can actually use is the hosted model. The open part is a promise, not a download.
Hy4 is Tencent's flagship Hunyuan model, a 770-billion-parameter mixture-of-experts system. Mixture-of-experts means only a fraction of those parameters activates for any given token, which is how a model of this size stays servable at all — the headline number is total capacity, not per-token compute, and it should not be compared directly with a dense model's parameter count.
Two capabilities are new to the Hunyuan line, and both are worth understanding separately:
Around those two, Tencent reports stronger instruction following, better long-document understanding, and longer-horizon agent task execution. The context window is listed at 1,048,576 tokens on OpenRouter's listing — a million-token window, in the same class as the largest frontier models. That figure comes from a gateway listing, not from Tencent's own announcement materials quoted in the coverage, so treat it as a listing figure rather than a vendor-confirmed spec.
Here is the distinction this article turns on:
Tencent's own framing — labelled as Tencent's, not as an independent finding — is that this is the first model in the Hunyuan line to combine visual understanding and computer use, and that its line-up is one of the strongest open-weight models available today. Tencent also admits, in its own materials, that Hy4 trails leading closed models on research maths, a point the coverage preserves rather than smooths over.
One independent data point, also labelled: the South China Morning Post reports that the Hy4 preview achieved a 23rd-place ranking on WebDev, a global AI benchmark leaderboard — a ranking for the preview, not for Hy4 itself.
Pricing for a model like this is where coverage most often goes wrong, by mixing a vendor price, a gateway price, and a promotion into a single number. Here they are kept separate:
No other pricing has been invented here. If you are comparing Hy4 against another model on cost, compare like with like: vendor price to vendor price, gateway price to gateway price, and promotional price only within its stated window.
Three things make Hy4 worth attention even before the weights arrive.
First, vision plus computer use in one flagship model from a major lab is a statement about where the frontier is heading: models that do not just generate text, but perceive screens and act on machines. That is the direction agent builders have been waiting for, and a major Chinese lab shipping it at flagship scale changes the competitive conversation.
Second, the open-weights promise matters. A 770-billion-parameter model is a data-centre-scale artefact, and the gap between "announced as open-weight" and "weights actually published" is where many launches quietly stall, so watch for the licence and the actual download.
Third, the competitive context is real. The coverage frames Hy4 against an intensifying open-weight race in which DeepSeek, Qwen, Kimi and others are shipping large, capable models — and against the established closed frontier. Hy4's reception will be shaped by that comparison, not by Tencent's framing alone.
We have not tested Hy4 hands-on, and nothing here is a hands-on finding. Everything above comes from the 4 November 2026 coverage cited at the end, and the limits of that coverage are the limits of this article.
No hands-on testing was performed for this article, and no benchmarks or prices have been invented. The bottom line is a measured one. Hy4 is a big, serious release: a 770-billion-parameter model that can see and can use a computer, available now as a hosted service, with its open weights still to come. What matters next is not the announcement, but the download — and whether the capabilities hold up when people outside Tencent finally get to test them.
Sources: South China Morning Post, 4 November 2026, “Tencent unveils Hy4” (https://www.scmp.com/tech/big-tech/article/3373227/tencent-unveils-hy4-open-weight-ai-race-deepseek-); PANews (https://www.panews.io/articles/01a1a509-67cb-767f-b692-bac8d93c7f9e); ChainCatcher (https://www.chaincatcher.com/en/article/2301648); Wedoany (https://en.wedoany.com/shortnews/504520.html); Longbridge (https://longbridge.com/en/news/304230896). Tencent's framing and the SCMP WebDev ranking are labelled as such in the article. No hands-on testing was performed.
What is Tencent's Hy4?
Hy4 is Tencent's new-generation Hunyuan large language model, released 4 November 2026 as the successor to the Hy4 preview.
What is new in Hy4 compared with earlier Hunyuan models?
Two firsts for the Hunyuan line: visual understanding (fine-grained image processing, complex chart analysis, mathematical visual reasoning) and Computer Use (operating a computer via its graphical interface and command line).
Can I download Hy4's weights today?
No. The weights are not yet released. Tencent says they will be open-sourced soon, but no date has been given.
How much does Hy4 cost?
Tencent Cloud preview-level pricing is RMB 6 per million input tokens, RMB 18 per million output tokens, and RMB 0.3 per million cache-hit tokens. TokenHub is 50% off until 31 December 2026, and WorkBuddy/CodeBuddy is free from 4 to 15 November 2026. OpenRouter lists $0.834 per million input tokens and $2.501 per million output tokens as a gateway price.
What is Hy4 Flash?
A lighter version of Hy4 focused on speed and cost, announced as coming soon. No specs, price or release date have been given.
How good is Hy4, honestly?
Tencent's own framing calls its line-up one of the strongest open-weight models available today, while also admitting Hy4 trails leading closed models on research maths. No independent hands-on results exist yet.
What is the context window?
1,048,576 tokens, per OpenRouter's listing. Treat that as a gateway listing figure rather than a vendor-confirmed specification.
Was Hy4 tested hands-on for this article?
No. No hands-on testing was performed; all facts come from the 4 November 2026 coverage cited above, and no benchmarks or prices have been invented.