Tencent released Hy4 on 4 November 2026, the full version of the model it previewed in August. It can read images and operate a computer for the first time, and is available through Tencent products and hosted APIs. Its weights, however, are not yet downloadable.

Tencent says Hy4 now ranks among the strongest open models. That claim needs a qualifier: the company plans to release the weights and a fuller system card, but launch-day checks found no final weights repository and no firm date in the launch post. Until those files appear, Hy4 is a hosted model, not one developers can run on their own servers.

Key takeaways

  • Hy4 adds vision and computer use, including command-line and graphical interface control, for the first time in the Hy line.
  • It is live in Yuanbao, WorkBuddy, CodeBuddy, ima and WeCom’s Dayuan assistant, and through Tencent Cloud TokenHub and OpenRouter.
  • API list prices are unchanged from the preview, with a TokenHub discount through 31 December 2026.
  • Tencent’s own benchmarks show strength in office and finance tasks, but a wide gap to leading closed models on terminal work and frontier mathematics.
  • Weights, a full system card and a cheaper Hy4 Flash version are promised, not yet delivered.

What Tencent actually released on November 4

Hy4 succeeds Hy4 Preview, released on 28 August 2026. Tencent says it substantially scaled up pre-training and post-training, with gains in instruction following, long-document understanding, long-horizon tasks and deep research. The launch was announced from Shenzhen through a company release carried by PRNewswire, and Reuters reported the same day that its cybersecurity training was limited to defensive content.

The hosted-versus-open distinction is the key detail. The preview was open-sourced on arrival in August, but the full version’s weights are described only as planned, with support promised from the vLLM and SGLang communities. A launch-day check found only preview repositories on Tencent’s Hugging Face account, and the weights link on OpenRouter did not lead to a publicly accessible file. The company release says weights and system cards will arrive in the coming days; the launch post itself gives no date, as translated by independent coverage.

Tencent Hy4 vision and computer use: what is new

Vision is the first headline addition. Hy4 can now take image input, read fine-grained detail and analyse complex charts, and combine that visual understanding with reasoning and tool use. In everyday work, that means reading documents, slides, diagrams and screenshots as part of a task, rather than needing text alone.

Computer use is the larger change. Hy4 can work through both a command line and a graphical interface, with dedicated training on professional software such as Blender and computer-aided design tools, and on long tasks spanning several applications. Tencent says Hy4 can also edit Word, Excel and PowerPoint files alongside a user in real time, and that game-development scores improved by 19.1 percent on GameDev and 9.8 percent on GameLogic compared with the preview.

Tencent Hy4 price and where to use it

Hy4 is available inside Tencent’s own products: the Yuanbao chatbot, WorkBuddy, CodeBuddy, the ima knowledge tool and WeCom’s AI Assistant, known as Dayuan. Developers can also call it through Tencent Cloud TokenHub and third-party API platforms including OpenRouter, Hermes and OpenCode. OpenRouter lists a context window of 1,048,576 tokens, with Tencent Cloud as the only listed provider there.

Pricing is unchanged from the preview: 6 yuan per million input tokens, 18 yuan per million output tokens and 0.3 yuan per million cached input tokens, about 0.83 dollars, 2.50 dollars and 0.04 dollars respectively. Tencent’s release rounds those to approximately 0.90 dollars, 2.70 dollars and 0.045 dollars. OpenRouter currently discounts the list prices by 50 percent, to about 0.42 dollars and 1.25 dollars. WorkBuddy and CodeBuddy are free from 4 to 15 November 2026, and TokenHub is discounted by 50 percent through 31 December 2026. A cheaper Hy4 Flash version is scheduled soon, with no price or date published in the material reviewed here.

Tencent Hy4 benchmarks: office strength, terminal weakness

Tencent’s strongest numbers come from workplace evaluations it ran itself. In office work on WorkBuddy, Hy4 averaged 83.18 with a pass rate of 96.15 percent, ranking first in nine of twelve scenarios in the mainland Chinese market. In finance, 213 professionals, including 153 experts from external institutions, judged 579 real-world tasks. Hy4 scored 3.04 out of 4, ahead of DeepSeek-V4.1-Flash at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 each. No closed model appears in that comparison, so it shows leadership among open models on that test, not overall.

The broader table is more mixed. Tencent published 34 benchmarks against its preview, GLM-5.3, DeepSeek-V4.1-Flash and Kimi K3.1 Pro, plus closed models GPT-6.1 Sol and Claude Opus 5.5. Twenty are Tencent’s own internal tests and most competitor scores are Tencent’s own runs, so this is a company presentation, not an independent league table. Even on Tencent’s numbers, Hy4 leads on only two benchmarks: APEX-Agents v1.1 and the internal Hy-LifeSearch test.

Selected Tencent benchmark scores for Hy4, published 4 November 2026. All figures are Tencent’s own runs.
Benchmark Hy4 Claude Opus 5.5 Kimi K3.1 Pro GPT-6.1 Sol
OSWorld 2.1, computer use 78.9 81.8 79.6 75.2
Terminal-Bench 4.0, terminal work 44.4 66.4 66.8 55.5
ProgramBench, coding 30.0 56.5 Not reported here Not reported here
Hy-FrontierMath T5, internal mathematics 13.0 91.3 Not reported here 97.5
ProofBench, internal test 5.5 42.2 Not reported here 70.5

The computer-use result is close: 78.9 on OSWorld 2.1 against 81.8 for Claude Opus 5.5 and ahead of GPT-6.1 Sol at 75.2. The terminal result is not: 44.4 against 66.4 and 66.8 for Claude and Kimi. The internal mathematics and proof tests show an even wider gap, which Tencent acknowledges. One caveat: Claude Opus 5.5 was blocked by content filters on 41 ProgramBench instances in Tencent’s run, which may depress that score. Until independent evaluators repeat these tests, Hy4 looks competitive for computer use and office work, and clearly behind on the hardest technical tests Tencent published.

The preview’s specifications, 770 billion total parameters with 49 billion active and a one-million-token context window, are widely repeated in launch coverage. Tencent has not published the final version’s architecture or licence in the material reviewed here, so treat those as preview figures.

Tencent Hy4 safety: defensive cybersecurity and biosafety limits

Safety is the second half of the launch story, and the reason Reuters led with it. Tencent says Hy4 was trained on defensive cybersecurity data, prioritising proof-of-concept generation and vulnerability remediation, while offensive material related to exploitation was excluded. Reuters summarises the filtering as removing exploit development, reverse engineering and penetration-testing content. Tencent’s Xuanwu and Keen labs assessed related areas and report no significant concerns in the areas evaluated.

That framing arrives when AI-driven cyberattacks are a live concern. The United Kingdom’s National Cyber Security Centre warned only the previous day that AI-powered ransomware was highly likely to hit British critical infrastructure by 2028, as covered in our NCSC AI ransomware warning. Tencent’s approach also follows Moonshot AI limiting the cyber-offensive capabilities of its newest model last month, as Reuters notes. On biosafety, Tencent reports 30.4 out of 100 on its internal Hy-BioExperimentBench test, below its own high-risk threshold. These are internal assessments, not independent audits.

What is still missing, and what it means

Three promised items define what this release is not yet: the model weights, the comprehensive system card, and Hy4 Flash. Until the weights appear, no team can self-host Hy4 or inspect its licence terms for the final version. Every request goes through Tencent Cloud, the only provider listed on OpenRouter, which also notes that EU in-region routing is not currently available. By contrast, Mistral released the weights of its large Mistral Large 4 model on 29 October, as covered in our Mistral Large 4 open-weights guide.

For users inside Tencent’s ecosystem, Hy4 is immediately useful: a chatbot and office assistant that can read images, work across documents and applications, and handle longer tasks. For developers, the choice is narrower. For office productivity, finance research or computer use through an API, Tencent’s evidence puts Hy4 among the strongest open models. For terminal-heavy engineering, frontier mathematics or private self-hosting, the evidence points elsewhere: to Claude Opus 5.5 or Kimi K3.1 Pro on terminal work, as detailed in our Kimi K3.1 Pro guide, or to a model whose weights are already available. Model prices have also fallen sharply this year, as covered in our guide to the 80 percent crash in AI model prices, so Hy4’s unchanged price is a holding position in a falling market, not a price cut.

Frequently asked questions

Is Tencent Hy4 open source yet?

Not in the practical sense. Tencent calls Hy4 an open-source model and says it plans to release the weights, but on 4 November 2026 the weights were not downloadable and no firm date had been published in the launch post.

What is new in Hy4 compared with Hy4 Preview?

Vision and computer use are new. Hy4 can read images and complex charts, and can operate a computer through command-line and graphical interfaces.

Conclusion

Hy4 is a significant step for Tencent: the first Hy model that can see and use a computer, offered at an unchanged price with a clear office and finance focus. Its limits are equally clear in Tencent’s own material, from the terminal and mathematics gaps to the safety tests awaiting a full system card. The release to watch next is not another announcement, but a file: the model weights. When they land, with a licence and a system card, Hy4 can be judged as the open model Tencent describes.

Sources and Further Reading