Tencent's Hy4 Model Adds Vision and Computer Use; Its Weights Are Still Closed

Tencent's Hunyuan team released the full version of its Hy4 large language model on 4 November 2026, adding image understanding and the ability to operate command-line and graphical software for the first time in the Hy line, according to a report on the launch by ChainCatcher. The release follows an Hy4 preview that Tencent had already published as an open-weight model on Hugging Face in August 2026.

The full version keeps the preview's architecture: a mixture-of-experts design with 770 billion total parameters, 49 billion of them active per token, and a context window of 1 million tokens, according to Tencent's own model card for Hy4-preview on Hugging Face. ChainCatcher reports that Tencent expanded the pre-training and post-training behind the final version and added vision understanding plus "computer use" — control of both terminal commands and graphical interfaces — for the first time in the series. Dowson Tong, Tencent's Cloud and Smart Industries Group CEO, said ahead of the launch that the update "focuses on strengthening long-text processing and various real-world productivity tasks, with markedly optimized chain-of-thought reasoning efficiency and vision understanding added for the first time," according to BigGo Finance's report on the announcement.

For developers and researchers comparing open and closed AI systems, what matters most is not the headline capability list but what Tencent's own numbers show, what an outside tester found, and what still has not shipped.

Tencent's own test results

At launch, Tencent reported that Hy4 scored 3.04 on a four-point scale in a financial-analysis evaluation covering 579 tasks and judged by 213 in-house experts, ahead of Kimi K3.1 Pro and GLM-5.3 (2.68 each) and DeepSeek-V4.1-Flash-1017 (2.97), per ChainCatcher's report. On the office-productivity benchmark Tencent built for its own WorkBuddy assistant, Hy4 reached an average score of 83.18 out of 100 and a 96.15% task pass rate, the same report says.

Tencent's own framing of those results was notably restrained: the team acknowledged "a significant gap compared to top proprietary models" while describing Hy4 as sitting among "the world's top-tier open-source models," ChainCatcher reports.

What independent testing shows

The benchmarking firm Vals AI evaluated Hy4 on 4 November 2026, the day it launched, running it through its own test suites rather than Tencent's internal evaluations. Vals AI put Hy4 at 59.35% on its general Vals Index, ranking it 13th of 34 models it has tested, and at 39.39% on Terminal-Bench 4.0, ranking it 15th of 61. On Vals AI's finance and legal indices, Hy4 scored 58.05% (13th of 37) and 70.55% (15th of 35) respectively. Those rankings place Hy4 in the middle of the field Vals AI tracks, a different picture from the head-to-head wins against named rivals that Tencent highlighted in its own evaluation.

The open-weight promise is not kept yet

Hy4's predecessor is already downloadable: the August preview sits on Hugging Face under an Apache 2.0 license, with Tencent's own model card listing scores including 92.3 on GPQA Diamond and 65.7 on SWE-Bench Pro. The November release is different — it ships only through Tencent Cloud, OpenRouter and Tencent's own apps such as Yuanbao and CodeBuddy. ChainCatcher reports that Tencent says the Hy4 weights and system card "will be open-sourced soon," without giving a date. No Hy4 repository distinct from the August preview had appeared on Hugging Face as of this writing.

What it means for developers and researchers

Hy4 is usable immediately through Tencent's API. ChainCatcher reports pricing of 6 yuan per million input tokens and 18 yuan per million output tokens for the new release — about $0.834 and $2.501 per million tokens, the same dollar figures Tencent Cloud's international documentation lists for the August preview, suggesting the per-token price has not changed. That gives developers a long-context, vision-capable model with agentic file and application control to test today. Anyone who wants to self-host it, inspect its weights, or fine-tune it offline still has only the smaller-scope August preview to work with until Tencent publishes the final model. For students and researchers tracking claims about model capability more generally, the gap between Tencent's own benchmark wins and Vals AI's mid-table independent rankings is a reminder to check who ran a given test before taking a benchmark score at face value.