Tencent's flagship model can now read a screen and operate it, but the company is keeping the weights to itself for now, a reversal of its usual open playbook.
Tencent switched on vision and computer-use capabilities for Hunyuan Hy4 on November 4. Ten weeks earlier, the model couldn't see an image at all. The preview version released August 28 was text-only, a 770-billion-parameter mixture-of-experts system built for coding and office work. This full release can interpret charts and documents, reason across multiple images, and operate command-line tools and graphical interfaces on its own. That puts it in the same small category as Anthropic's and OpenAI's computer-use agents.
Anyone hoping to run it themselves can't, not yet. Tencent shipped Hy4 as API-only and said the model weights and system card would be released later. That's notable because the August preview was released under an Apache 2.0 license, with downloadable weights on Hugging Face. It was part of the same open playbook DeepSeek, Alibaba's Qwen and Moonshot's Kimi have used all year to win developer mindshare. Pulling the weights back, even temporarily, breaks that pattern right as rivals keep shipping open checkpoints.
You can reach Hy4 through Tencent Cloud's TokenHub, through OpenRouter, or inside Tencent's own WorkBuddy and CodeBuddy products. In those products the company is running a free usage quota from November 4 through November 15. After that, it's priced at $0.834 per million input tokens and $2.501 per million output tokens. TokenHub also carries a 50% discount through the end of December. That's still a fraction of what Anthropic charges for Claude Opus.
Tencent didn't build this for general web browsing. The training focused on office tasks, financial analysis, and game development, including operating software like Blender and CAD tools. In those areas a model that can see a screen and manipulate it directly saves a human from describing every click. On an internal finance evaluation covering 579 real-world tasks, Tencent says Hy4 scored 3.04 against DeepSeek-V4.1-Flash's 2.97. The company also reported gains of 19.1% on its GameDev benchmark and 9.8% on GameLogic.
Tencent's Hy3 Bets on Smaller Agent Models Instead of Bigger Ones Tencent's Hunyuan team released Hy3 on July 6, a 295 billion parameter Mixture-of-Experts model that activates only 21 billion parameters and is built specifically for agent work rather than benchmark size. Internal WorkBuddy tests show task success jumping from 72% to 90% and hallucinations dropping from 12.5% to 5.4%, part of a wider trend of... - smaller language models for AI agents - Tencent WeChat AI agent strategy
The more interesting number is hallucinations. Tencent says its internal search evaluations showed Hy4's hallucination rate drop from 15.76% in the August preview to 5.45% in this release. If that holds up outside Tencent's own test harness, it matters more for agentic use than any leaderboard score. A computer-use agent that occasionally invents a file path or misreads a dialog box is a liability, not a convenience.
Tencent did publish a head-to-head computer-use benchmark, though the result is company-reported rather than independently reproduced. Hy4 scored 78.9 on OSWorld 2.1, compared with 81.8 for Claude Opus 5.5, 79.6 for Kimi K3.1 Pro and 75.2 for GPT-6.1 Sol in Tencent's table. The company also has a separate screen-operating model: UI-Mate-27B, released in August, scored 77.0 on OSWorld-Verified, ahead of Kimi-K2.6's 73.1 in its published evaluation.
A crowded field, and Tencent is late to the vision party #
Frankly, Tencent is playing catch-up here. DeepSeek, Alibaba's Qwen line and Moonshot's Kimi have all shipped multimodal, agent-capable open models through most of 2026, and Meituan's LongCat has pushed into the same territory. Tencent said total calls to Hy3 during its first week were more than 68 times those of the previous-generation Hy2. At its Global Digital Ecosystem Summit on October 29, Tencent previewed visual understanding for Hy4 and announced that Hy4 Flash was also coming.
But usage inside a billion-user ecosystem like WeChat and QQ is a different thing from winning developers who choose a model purely on capability. Those developers have spent 2026 picking among DeepSeek V4, Qwen 3.6, Kimi K3 and GLM-5.2, all open-weight, all competing hard on agentic coding. Hy4's decision to go API-only at launch, even if the weights do follow later as promised, costs it exactly the audience that made Hy3's download numbers worth bragging about.
For builders evaluating agentic tools right now, the honest read is this: Hy4 is inexpensive and Tencent has published computer-use results against competing models, but those scores have not yet been independently reproduced. Until the weights land and someone outside Tencent runs OSWorld on this exact model, treat the comparisons with Claude and OpenAI's computer-use agents as vendor-reported results. Also read: A free stealth AI model called Crouton Alpha just matched Gemini's context window • AI Infrastructure Stocks Sink as Democratic Midterm Sweep Threatens Data Centers • AI Is Quietly Ending the Fiverr and Upwork Era of Gig Work
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.