Tencent has just shipped the biggest upgrade in its Hunyuan AI line yet. Hy4, released on 4 November 2026, is a 770-billion-parameter model that does two things no Tencent model has done before: it can see images, and it can operate a computer. It is also, by Tencent's own admission, still a step behind the best closed models — which makes it an honest release to unpack.
Here is what Hy4 is, what it costs, where it beats its rivals, and whether you should care.
In This Article
What Is Tencent Hy4?
Hy4 is Tencent's flagship large language model, the successor to the Hy4 preview that launched on 28 August 2026. It is a mixture-of-experts (MoE) model with 770 billion total parameters, of which 49 billion are active per token according to the preview's model card, and a 1-million-token context window — enough to hold a whole codebase or a book-length document in a single prompt.
The headline is not the size. It is the two new senses: vision and computer use. Hy4 is the first model in Tencent's Hy series that can read images and operate software, pitched at professional work — office productivity, financial analysis, software development, and scientific research.
One important caveat up front: the full model's weights are not public yet. Tencent says it plans to release them "in the coming days", alongside a system card with fuller safety results. The August preview's weights are already on Hugging Face under the Apache 2.0 licence. Until the final weights land, Hy4 is a model you reach through Tencent's apps and APIs — not one you download.
What's New: Vision and Computer Use
It can see now
Hy4 adds visual understanding for the first time in the Hy line: fine-grained image processing, complex chart analysis, and mathematical visual reasoning. In one demo cited in the launch coverage, the model located a roof fitting about two metres across inside a 2.8-billion-pixel aerial photograph.
It can use a computer
The bigger change is computer use. Hy4 can work through both the command line and graphical interfaces: reading on-screen text, clicking buttons, and completing multi-step tasks across applications — scheduling appointments, managing spreadsheet data, and routing emails. Tencent says it trained the model specifically on professional software such as Blender and CAD tools, and on long tasks that span several apps.
This puts Hy4 in the same conversation as the other computer-use agents we have covered — OpenAI's GPT-6.1 Terra and the open-source NorthCUA — except Hy4 arrives as a single giant model rather than a small specialist agent.
Tencent also reports its internal hallucination rate fell from 15.76% to 5.45% versus the preview — its own figure, on an unpublished test set, so treat it as a direction of travel rather than a verified fact.
Hy4 Benchmarks: The Honest Read
Every benchmark number below comes from Tencent's own testing, in which it re-ran rival models under consistent settings. None of it is an independent leaderboard — and to Tencent's credit, its own chart shows Hy4 trailing the frontier on most tasks.
| Benchmark | Hy4 | Claude Opus 5.5 | GPT-6.1 Sol |
|---|---|---|---|
| Terminal-Bench 4.0 (coding agents) | 44.4 | 66.4 | 55.5 |
| OSWorld 2.1 (operating a computer) | 78.9 | 81.8 | 75.2 |
| E-Bench 2.0 | 70.0 (2nd of 7) | 80.0 | — |
The pattern is clear. On raw coding-agent work (Terminal-Bench), Hy4 sits well behind Claude Opus 5.5. On actually operating a computer (OSWorld), it is nearly level with Claude and ahead of GPT-6.1 Sol — as expected, given its computer-use training. On E-Bench 2.0 it ranks second of the seven models Tencent compared. For the wider coding-assistant field, see our Best AI Coding Assistants in 2026 roundup.
Against other open-weight models, the picture is mixed. In a blind evaluation for financial work run with Tencent's fintech unit — 213 financial professionals rating outputs on 579 real tasks — Hy4 averaged 3.04 out of 4, ahead of DeepSeek V4.1 Flash (2.97) and Kimi K3.1 Pro and GLM-5.3 (2.68 each). But across Tencent's wider 12-benchmark table, Hy4 reportedly ranks first on none of them, and wins six of twelve head-to-heads against Kimi K3.1 Pro while losing the other six.
In short: Hy4 is a strong open-weight contender that trades blows with its Chinese open rivals, but it is not yet a frontier-beater. Tencent says so itself — the launch materials describe Hy4 as ranking "among the world's most capable open-weight models" while acknowledging a considerable gap with the leading closed models.
Hy4 Pricing: Cheaper Than the Frontier
Where Hy4 genuinely stands out is price. Tencent kept the preview's pricing:
- Input: 6 yuan (about $0.83–$0.90) per million tokens
- Output: 18 yuan (about $2.50) per million tokens
- Cached input: 0.3 yuan (about $0.04) per million tokens
TokenHub charges 50% off those list prices through 31 December 2026, and OpenRouter lists Hy4 even lower, at $0.42 per million input tokens and $1.25 per million output tokens.
For comparison, Gemini 4 Carbon costs $4/$20 per million tokens. Hy4 is roughly a fifth of that on input. A cheaper, faster variant — Hy4 Flash — is due soon, aimed at maximum price-performance for high-volume tasks.
How to Try Hy4
Hy4 is live now in Tencent's own products: TokenHub, WorkBuddy, CodeBuddy, the Yuanbao chatbot, the WeCom assistant Dayuan, and ima. Outside China, developers can call it through OpenRouter (tencent/hy4), Hermes, and OpenCode. WorkBuddy and CodeBuddy are offering a limited free quota from 4 to 15 November, so that is the cheapest way to test it hands-on.
If you want the weights, you will have to wait — Tencent has promised them "shortly" with support from the vLLM and SGLang projects, but gave no date, and has not said which licence the final weights will carry.
One practical warning: do not plan to run Hy4 on your own hardware. The preview's weights need roughly 924 GB even quantised (1,848 GB in full BF16 precision), so this is a data-centre or API model, not a laptop one.
Hy4 vs Its Rivals
Hy4 lands in a crowded open-model field:
- vs Kimi K3.1 Pro: The closest fight. Tencent's own table has them tied 6–6 across twelve benchmarks, though Hy4 won the finance blind evaluation. Kimi is better known in the West; Hy4 is cheaper and now has computer use. See our Kimi K3.1 Pro explainer.
- vs DeepSeek V4.1 Flash: Hy4 edges it in Tencent's finance evaluation (3.04 vs 2.97) and adds vision and computer use.
- vs Claude Opus 5.5: Not a fair fight yet. Claude leads clearly on coding (66.4 vs 44.4 on Terminal-Bench), though Hy4 nearly matches it on computer use — at several times lower cost.
- vs GPT-6.1 Sol: Hy4 trails on coding but beats Sol on OSWorld computer-use scores, at a fraction of the price.
For a wider view of the open-model leaderboard, see our comparison DeepSeek vs Qwen vs Kimi: Best Open AI Model in 2026.
Pros and Cons
Pros:
- Vision + computer use in one open-weight model — a first for Tencent's line
- Very cheap: about $0.83/$2.50 per million tokens, half price until 31 December
- 1M-token context window for huge documents and codebases
- Near-frontier scores on computer-use benchmarks (OSWorld 78.9)
- Available now via OpenRouter and Tencent's apps, with a free trial window
Cons:
- Full weights not yet released — "open-weight" is a promise, not a download, for now
- Trails Claude and GPT-6.1 Sol clearly on coding benchmarks
- All benchmark figures are Tencent-reported, not independently verified
- Far too large to self-host without serious GPU hardware
- No confirmed licence for the final weights yet
Verdict: Who Is Hy4 For?
Hy4 is best understood as a value play, not a leaderboard play. If you need the absolute best coding agent, Claude Opus 5.5 still wins — and Tencent's own numbers say so. But if you need a cheap, capable model that can look at a screen, read a chart, and operate software across a long task, Hy4 is now one of the most interesting options available, at roughly a fifth of the frontier price.
It will matter most to three groups: developers building agents on a budget, businesses that want computer-use automation without frontier-model bills, and teams in Tencent's ecosystem (WeChat Work, Yuanbao, CodeBuddy) where Hy4 is now the default brain. The open-weights release, when it lands, will be the real test.
FAQ
Is Tencent Hy4 open source?
Not exactly. The August preview's weights are published under the Apache 2.0 licence, but the full Hy4 weights have not been released yet — Tencent says they are coming "in the coming days". Open-weight is the plan; the download is still pending.
How much does Hy4 cost?
About $0.83 per million input tokens and $2.50 per million output tokens on Tencent Cloud TokenHub, with 50% off through 31 December 2026. OpenRouter lists it even cheaper, at $0.42/$1.25 per million tokens.
Can Hy4 really use a computer?
According to Tencent, yes — it can operate command-line and graphical interfaces and complete multi-step tasks across apps. On Tencent's own OSWorld 2.1 test it scored 78.9, close to Claude Opus 5.5's 81.8.
Is Hy4 better than Claude or GPT-6.1?
On coding benchmarks, no — it trails both clearly on Tencent's own figures. On computer-use tasks it is nearly level with Claude and ahead of GPT-6.1 Sol, at a much lower price. It is a value pick, not a frontier-beater.
Sources: Emergent, DepthGrid, WinBuzzer, MadRobot. All benchmark figures are Tencent-reported.
No comments:
Post a Comment