I checked Tencent's own announcement this morning, then cross-checked it against two independent reports, before writing this — because "open-source" gets thrown around loosely on launch day, and I wanted to know exactly what is live today and what is still a promise.
What actually launched today
Tencent has released the full version of Hy4, its latest Hunyuan large language model, on November 4, 2026. It follows the Hy4 preview that arrived in August 2026. Tencent describes this as a significant step up in performance, and says the model now ranks among the world's leading open-source models. That ranking claim is Tencent's own; independent benchmarks for the full version will take time to appear, and I would wait for them before repeating any leaderboard claim as fact.
Vision and computer use, explained simply
Two things are genuinely new here. First, vision: Hy4 can now interpret fine-grained image details, complex charts, and mathematical problems presented as images. Second, computer use: it can operate across command-line and graphical interfaces. Put together, that means the model is meant to look at a screen, understand what is on it, and then take steps to get a job done — rather than only answering a question in text. Tencent points to office collaboration, financial analysis, software development, and tasks that span several applications as the places this should show up first.
Where you can try it right now
Hy4 is available through hosted routes today: Tencent Cloud TokenHub and OpenRouter, and inside Tencent products including WorkBuddy, CodeBuddy, Yuanbao, ima, and the WeCom assistant Dayuan. API pricing is unchanged from the preview, at roughly 0.90 dollars per million input tokens and 2.70 dollars per million output tokens, with cached input at about 0.045 dollars per million. WorkBuddy and CodeBuddy are free from November 4 to 15, and Hy4 is 50 percent off on Tencent Cloud TokenHub through December 31, 2026. A cheaper Hy4 Flash version is promised soon, aimed at faster, lower-cost inference.
The catch: the weights are not out yet
Here is the part worth reading twice. The earlier preview was open-sourced, but for this full version Tencent says it plans to release the model weights and a comprehensive system card in the coming days, with ecosystem support from the vLLM and SGLang communities. Until those files actually appear, this is a hosted release you access through Tencent's services and partners — not yet a model you can download and run yourself. That distinction matters, because it is the difference between "open-source today" and "open-source soon." A similar pattern has shaped other recent open-weight launches: the announcement lands first, and the weights follow. Reflection AI's Beam launch, for example, arrived with its own staggered release plan, and Moonshot's Kimi K3.1 Pro and Mistral's Le Chonk each had to be judged the same way — on what was actually available on day one, not only on what was promised.
Why this matters — our take
My verdict: the vision-plus-computer-use combination is the real upgrade, because it moves Hy4 from answering questions toward doing work across apps. That is useful for anyone who lives in documents, spreadsheets, and code. The honest limit is the safety story. Tencent says it trained Hy4 with limited cybersecurity data, filtered to defensive content, filtering out material such as exploit development and penetration testing, to reduce the risk of the model helping with cyberattacks. That is a deliberate, welcome choice — but it is also a trade-off, and Tencent itself has acknowledged a gap remains between Hy4 and the top closed models. Who should care: developers and office teams already inside Tencent's ecosystem, and anyone tracking how fast Chinese open models are closing the gap. Who should skip the hype: anyone who needs a downloadable model today, or a fully independent safety evaluation — neither is in hand yet.
What to watch next
Three things will decide whether this launch ages well. First, do the weights and system card actually arrive in the coming days, as promised? Second, do independent tests confirm the vision and computer-use gains outside Tencent's own evaluations? Third, does Hy4 Flash deliver the promised cost cut without losing the qualities that make the full model useful? I will be watching the weights release first — that is the moment this stops being a hosted announcement and becomes the open model it is being presented as.
What do you think — is a model that can see your screen and click through tasks a genuine productivity leap, or a step too far for trust? Tell us in the comments.
Sources
- Tencent official announcement: https://www.tencent.com/tencent-launches-hy4-joining-the-ranks-of-the-worlds-leading-open-source-ai-models/
- Channel News Asia report on Hy4's limited cybersecurity training data: https://www.channelnewsasia.com/business/tencent-rolls-out-ai-model-trained-limited-cybersecurity-data-amid-hacking-fears-6523736
- Unite.AI report on the full Hy4 release: https://www.unite.ai/tencent-releases-full-hy4-model-with-vision-and-computer-use/


Comments
Post a Comment