Tencent Hy4 Is Here With Vision and Computer Use — but You Still Cannot Download It

Tencent's Hy4 is live — and its first day broke the queue

Tencent released the full version of Hy4, its flagship Hunyuan model, on 4 November 2026, adding vision and computer use for the first time in the Hy line, and demand was so strong that the company expanded the model's inference capacity tenfold within a day. Hy4 is a mixture-of-experts model, which means it routes each token to a small subset of specialist sub-models rather than running all 770 billion of its parameters at once, and it is available now only through Tencent's own services and APIs. The downloadable weights, which Tencent says will be open-sourced soon, are still not out.

Hy4 is the fourth generation of Tencent's Hunyuan family and the company's new flagship for agentic, productivity-focused work. It follows an Hy4 preview released in late August 2026, which was text-only and whose weights were published on Hugging Face under an Apache 2.0 licence. The November release is the full model, and it is different in two ways that matter: it can see images, and it can operate a computer. Users of WorkBuddy, Tencent's AI agent for office work, hit the free-usage quotas on Hy4's first day there, and Tencent said on 5 November that it had expanded the Hy4 inference cluster tenfold and raised the free usage quota.

Illustration created for INORYVA

What is actually new

Computer use is the headline feature, and it is worth defining plainly. Computer use means the model can interact with both command-line interfaces and graphical user interfaces to perform work on its own, rather than only answering in text. Hy4 was given specialised training on professional software such as Blender and CAD, which Tencent says significantly improves its ability to plan and execute long-horizon tasks. Vision is the other first: Hy4 supports visual understanding for the first time in the Hy line, including fine-grained image processing, complex chart analysis, and mathematical visual reasoning.

Behind the new features, Tencent says it substantially scaled up both pre-training and post-training compared with the preview. The headline size figures are carried over from the preview's own model card, because Tencent has not yet published a separate model card for the final version: 770 billion total parameters, with 49 billion active per token, and a context window of 1 million tokens. A lighter version, Hy4 Flash, is scheduled for release soon, but Tencent has given no specs, no date and no price for it. Hy4 is live across Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom's AI assistant Dayuan, and ima, with API access through OpenRouter as tencent/hy4, plus Hermes and OpenCode. From 4 to 15 November, WorkBuddy and CodeBuddy are available for free.

What it costs

API pricing is unchanged from the preview. Tencent Cloud's TokenHub lists 6 yuan per million input tokens, 18 yuan per million output tokens, and 0.3 yuan per million cached input tokens, which works out to about $0.83, $2.50 and $0.04 per million tokens. With the 50 percent TokenHub discount running until 31 December 2026, the effective TokenHub prices are half those figures until the offer ends. Tencent's own press release rounds the same prices to about $0.90 and $2.70, which is the same price expressed more roughly, not a different price. The catch is the one this article keeps returning to: you can only rent it. Until the weights ship, nobody outside Tencent can download Hy4, run it locally, fine-tune it, or inspect it.

Illustration created for INORYVA

Tencent's own scores vs independent tests

Tencent's launch materials are full of benchmark wins, and every one of them is Tencent's own test. The most eye-catching is a financial-analysis evaluation: in a blind test of 579 real-world financial tasks judged by 213 financial professionals, including 153 experts from outside institutions, Hy4 scored 3.04 on a four-point scale, ahead of DeepSeek-V4.1-Flash-1017 at 2.97 and both Kimi K3.1 Pro and GLM-5.3 at 2.68. On WorkBuddy's own office-productivity benchmark, Hy4 reached an average score of 83.18 out of 100 with a 96.15 percent task pass rate. In internal evaluations, Hy4 cut its hallucination rate from 12.95 percent in the preview to 7.2 percent. None of those results has been independently verified.

Vals AI, an independent benchmarking firm, evaluated Hy4 on launch day, 4 November, running it through its own test suites rather than Tencent's internal evaluations.

Test	Tencent's own figure	Vals AI independent figure
Terminal-Bench 4.0 (coding agents)	44.4 (Tencent's own test)	39.39 percent, 15th in Vals AI's table
Vals Index (general, finance, coding, legal, tax)	Not reported by Tencent	59.35 percent, 13th of 34 models
Vals finance index	3.04 of 4 in Tencent's blind finance test	58.05 percent, 13th of 37

Read the table as a lesson in who ran the test. On Tencent's own benchmarks, Hy4 beats named rivals, including Kimi K3.1 Pro. On Vals AI's independent suites, Hy4 sits in the middle of the field: 13th of 34 on the general index, 15th on Terminal-Bench 4.0. Vals AI's Terminal-Bench 4.0 score for Hy4, 39.39 percent, is below Tencent's own 44.4 on the same benchmark name, because the two sides ran it in different harnesses.

The comparison with Kimi K3.1 Pro is the one most readers will make. Our report on Kimi K3.1 Pro covered Moonshot's 2.8-trillion-parameter model, whose weights were actually published and downloadable. Hy4 is smaller on paper and cheaper through the API, but Kimi's weights are out and Hy4's are not. The other open-weight peer is Mistral Large 4, covered in our report on Mistral Large 4, which shipped with open weights from day one.

The weights question

Everything about this release comes back to the weights. Tencent's press release says the company plans to release Hy4's model weights and system cards "in the coming days", with a focus on cybersecurity and biological safety. The preview's weights are already on Hugging Face under Apache 2.0, and Tencent says it is working with vLLM and SGLang on support for the final model. But as of 7 November, no Hy4 repository distinct from the August preview had appeared on Hugging Face, and Tencent has given no date. The safety section of the announcement is part of the reason for the wait: Hy4 was trained on defensive security data, with offensive materials related to exploitation explicitly excluded, and experts from Tencent's Xuanwu Lab and Keen Security Lab found no significant concerns in the areas they evaluated. The system card that will come with the weights is supposed to carry the full safety evaluation.

Illustration created for INORYVA

Who should try it, and who should wait

If you build office-productivity tools, financial analysis workflows, or agentic software that operates desktop applications, Hy4 is worth a test run this week, while WorkBuddy and CodeBuddy are free until 15 November and TokenHub is half price until 31 December. Start with a small, well-defined task, watch for quota limits at peak hours, and keep Hy3 as a fallback inside WorkBuddy if Hy4 hits a limit. If you need to self-host, fine-tune, or inspect a model before putting it into production, wait. The August preview is the only Hy4 weight you can download today, and it is text-only, without vision or computer use. Do not describe Hy4 as open-source or open-weight in your own writing until Tencent publishes the weights.

For the wider market, Hy4 is another data point in the race among Chinese labs, and its price puts pressure on every other provider. The flagship comparison in our report on Gemini Carbon vs Claude Opus 5.5 vs GPT-6 Astra shows how tight the frontier race is, and Hy4's mid-table independent scores show how far the open-weight field still has to go to catch the proprietary leaders. Our coverage of Gemini 4 Carbon gives the other side of that comparison.

Key Takeaways

• Tencent released the full version of Hy4 on 4 November 2026, adding vision and computer use for the first time in the Hy line, after a text-only preview in August 2026.
• Hy4 is a 770-billion-parameter mixture-of-experts model with 49 billion parameters active per token and a 1-million-token context window, per the preview's model card.
• Demand on WorkBuddy forced a tenfold Hy4 capacity expansion on 5 November; free WorkBuddy and CodeBuddy access runs until 15 November 2026.
• API pricing is 6 yuan per million input tokens and 18 yuan per million output tokens, about $0.83 and $2.50, half price on TokenHub until 31 December 2026.
• Vals AI ranks Hy4 mid-table: 59.35 percent on the Vals Index, 13th of 34, and 39.39 percent on Terminal-Bench 4.0, 15th, below Tencent's own 44.4.
• The full model's weights are NOT released yet. Only the August preview weights are downloadable, under Apache 2.0.

Conclusion

Hy4 is a strong, cheap, genuinely useful model with two real firsts for Tencent: vision and computer use. It is also a release with an asterisk the size of the model itself. The weights are not out, the independent scores are mid-table, and the free access ends on 15 November. Try it through the API while it is free, and watch for the weights.

Sources

• Tencent — Tencent Launches Hy4, Joining the Ranks of the World's Leading Open-Source AI Models, 4 November 2026 (updated 6 November 2026), opened and read in full: https://www.tencent.com/tencent-launches-hy4-joining-the-ranks-of-the-worlds-leading-open-source-ai-models/
• Vals AI — Vals Index Leaderboard and Methodology, updated 5 November 2026: https://www.vals.ai/benchmarks/vals_index?suggested=pareto-efficient-plot
• Vals AI — Terminal-Bench 4.0 Leaderboard and Methodology, updated 5 November 2026: https://www.vals.ai/benchmarks/terminal-bench-4
• TechNode — Tencent expands Hy4 inference capacity tenfold as WorkBuddy demand surges, 6 November 2026: https://technode.com/2026/11/06/tencent-expands-hy4-inference-capacity-tenfold-as-workbuddy-demand-surges/
• EduFabTech — Tencent Hy4: Vision, Computer Use, Closed Weights, 6 November 2026: https://edufabtech.com/news/tencent-hy4-vision-computer-use-weights-not-open
• OmniKey — Tencent Hy4 vs Hy4 Preview: What Changed at Release: https://omniakey.com/blog/tencent-hy4