Tencent’s newest flagship model can now read images and operate a computer, and it costs the same as the preview it replaces. Tencent Hy4 launched early on Wednesday, November 4, at $0.834 per million input tokens and $2.501 per million output tokens, according to Tencent’s Hy4 research page. Tencent calls it one of the world’s leading open-source models.
That label is ahead of the facts for now. Hy4’s weights aren’t downloadable yet. Tencent says they’ll follow “in the coming days,” along with a system card. Until then, you can only reach Hy4 through Tencent’s API, OpenRouter and Tencent’s own apps. The preview it replaces shipped under Apache 2.0, and the open-weight field Tencent is measuring itself against (Kimi, DeepSeek, GLM) already lets you download the files.
- Price: $0.834 input, $2.501 output and $0.042 cached input per million tokens, unchanged from Hy4 preview. Tencent Cloud’s TokenHub halves that through December 31, 2026 (OpenRouter shows $0.417 and $1.251).
- Where: the API on Tencent Cloud and OpenRouter now, plus Tencent’s WorkBuddy and CodeBuddy (free November 4 to 15), Yuanbao and ima.
- Specs: 1,048,576-token context, 65,536-token max output, text and images in, text out. Tencent hasn’t published Hy4’s parameter count.
- The catch: no open weights, license or system card yet, and Tencent’s own chart has Hy4 well behind Kimi K3.1 Pro and Claude Opus 5.5 on Terminal Bench 4.0.
What Tencent Hy4 adds: vision and computer use
Hy4 is the first Tencent Hy model with vision, and the first trained for computer use, both at the command line and in graphical apps. Tencent says it trained the model on real software, naming Blender and CAD tools, and the launch video shows it building and lighting a room in Blender from a floor plan.
Tencent also says it worked with its game studios on front-end and game development. It reports gains of 19.1% on an internal GameDev test and 9.8% on GameLogic over the preview, and a 12.03-point jump to 61.63 on the IRAB agent benchmark. Those are Tencent’s numbers from Tencent-run tests. No outside lab has reproduced them yet.
Introducing Hy4 https://t.co/MdBwtx8m52 pic.twitter.com/4mcX5jiUDH
— Tencent Hy (@TencentHunyuan) November 4, 2026
Video: Tencent. “Introducing Hy4” from the official Tencent Hy account on X, November 4, 2026 (watch on X). It’s a 70-second edited launch reel, not a live test.
The price held, though Tencent’s two launch pages disagree
The research page and OpenRouter list the same per-million-token prices. Tencent’s press release rounds them to “approximately” $0.90, $2.70 and $0.045, so expect to see both sets quoted.
| Per million tokens | List price | TokenHub promo to Dec 31 |
|---|---|---|
| Input | $0.834 | $0.417 |
| Output | $2.501 | $1.251 |
| Cached input | $0.042 | $0.021 |
At list price, Hy4 costs about five times as much as DeepSeek V4.1 Flash 1017 on input and four times as much on output (we covered that update at $0.15 in and $0.60 out), though it’s still cheap next to closed frontier models. Tencent Cloud is the only provider OpenRouter lists, so the price is Tencent’s to set until other hosts can run the weights.
Tencent’s own chart shows where Hy4 trails
Tencent’s launch chart compares Hy4 with open and closed rivals, and it doesn’t flatter the model everywhere. On Terminal Bench 4.0, a test of agentic coding in a command line, Tencent scores Hy4 at 44.4. That beats GLM-5.3 (42.9) and DeepSeek-V4.1-Flash-1017 (40.0), but trails GPT-6.1 Sol (55.5), Claude Opus 5.5 (66.4) and Kimi K3.1 Pro (66.8). The gap is closer on OSWorld 2.1, a computer-use test, where Hy4 scores 78.9 against Kimi K3.1 Pro’s 79.6 and Opus 5.5’s 81.8.
Tencent’s strongest claim comes from a blind finance evaluation it ran with 213 finance professionals across 579 tasks. Hy4 averaged 3.04, ahead of DeepSeek-V4.1-Flash-1017 (2.97), Kimi K3.1 Pro (2.68) and GLM-5.3 (2.68). Tencent designed and ran that test.
The company is also candid about the ceiling. Its research page says “a notable gap remains” between Hy4 and top proprietary models in research-level math, long-horizon coding and advanced visual question answering. Even its hallucination numbers don’t match across pages: the research page says the rate fell from 15.76% to 5.45%, while the press release says 12.95% to 7.2%. Both put the drop in unsupported-assumption errors at 23.1% to 5.6%.
Open weights are promised, not shipped
As of 8:50 AM ET on November 4, Tencent’s Hugging Face page had no public Hy4 repository. Only Hy4 preview is there, posted on August 27 under Apache 2.0, with 770 billion total parameters and 49 billion active. Tencent hasn’t said whether the final model keeps that size or that license. It says the release will come with a system card and day-one support in the vLLM and SGLang inference engines.
Safety details are thin until that card arrives. Tencent says Tencent Xuanwu Lab and Keen Security Lab red-teamed the model, and it reports a 30.4 out of 100 score on its own biosecurity benchmark. Its pages disagree here too. The research page says reverse engineering and penetration-testing material was filtered out of training, while the press release says the model “incorporates” both for analysis. Tencent hasn’t explained the difference.
What to watch next
Three things settle whether Hy4 belongs in the open-weight conversation: the Hugging Face upload, the license attached to it, and a system card that shows how Tencent ran its benchmarks. Tencent also plans a smaller Hy4 Flash. Its press release headline says Flash was “recently released,” but the text says it’s “scheduled for release soon.” The free WorkBuddy and CodeBuddy period ends November 15, and the TokenHub discount runs to December 31.
Frequently asked questions
How much does Tencent Hy4 cost?
$0.834 per million input tokens, $2.501 per million output tokens and $0.042 per million cached input tokens, the same as Hy4 preview. Tencent Cloud’s TokenHub offers 50% off through December 31, 2026.
Can I download Tencent Hy4?
Not yet. Tencent says the weights and a system card will be released “in the coming days.” As of the morning of November 4, only Hy4 preview was on Hugging Face.
What is Hy4’s context window?
OpenRouter lists 1,048,576 tokens of context and up to 65,536 output tokens, with text and image input and text output.
Where can I use Hy4?
Through the Tencent Cloud API (TokenHub) and OpenRouter, and in Tencent’s WorkBuddy, CodeBuddy, Yuanbao and ima apps. WorkBuddy and CodeBuddy offer it free from November 4 to 15.
Sources
- Tencent Hy, “Hy4” research page, November 4, 2026: hy.tencent.ai/research/hy4
- Tencent press release, “Tencent Launches Hy4, Joining the Ranks of the World’s Leading Open-Source AI Models,” November 4, 2026: tencent.com
- Tencent Hy (@TencentHunyuan) launch post and benchmark chart on X, November 4, 2026: x.com/TencentHunyuan
- Tencent Hy, “Introducing Hy4” video on X, November 4, 2026: x.com/TencentHunyuan
- OpenRouter, Tencent Hy4 model page (pricing, context, providers): openrouter.ai/tencent/hy4
- Hugging Face, Hy4 preview model card: huggingface.co/tencent/Hy4-preview
- GitHub, Tencent-Hunyuan/Hy4-preview: github.com/Tencent-Hunyuan/Hy4-preview
All sources accessed November 4, 2026.
Last updated: November 4, 2026. To report an error, see our corrections page. Articles are drafted with AI assistance and reviewed and edited by an editor; see our editorial policy.
