Tencent Hy4 Launches: 770B Model With Vision and Computer Use, Weights Still Pending

Tencent Hunyuan released the full Hy4 on November 4, 2026. It is a 770-billion-parameter mixture-of-experts model with a 1 million token context window, and it is the first Hy-series model with vision and with computer use through both the command line and graphical interfaces. The earlier Hy4 preview from August 28 had neither.

The headline numbers come from Tencent's own tests. On those, Hy4 does not rank first on any of 12 benchmarks, and the official weights are listed as planned, not released. That makes this a strong, cheap model to try through an API, and a claim to check before you plan around it.

TL;DR: what readers ask first

What Tencent says changed since the preview

According to the Apiyi write-up of Tencent's release, Hy4 scales up both pretraining and post-training. The main gains are in instruction following, long-document comprehension, long-horizon tasks and deep research. Thinking mode is on by default, based on the preview documentation, and the API returns a reasoning_content field your code must handle. We could not load Tencent's own release page, which renders with JavaScript, so details here come from outlets and aggregators that quote it, and we mark vendor claims as such.

Vision. Tencent highlights fine-grained image handling, chart analysis and visual math. Its demos include locating a roughly 2-meter rooftop feature in a 2.8-billion-pixel aerial image and reassembling a 900-piece puzzle of a Monet painting.

Computer use. This is the larger change. Tencent says the model was trained for professional software such as Blender and CAD tools and for long tasks across applications. The Tencent Cloud Developer Community reportedly says the team built a sandbox with hundreds of installed applications to run tasks and check results. We covered the general idea in our look at agent sandbox design.

A cursor reaching toward a button, showing Hy4 computer use in a graphical interface

Games and frontend. Tencent says its game studio developers contributed to deep integration with Unreal Engine, Unity and Godot. Against the preview, it reports GameDev up 19.1% and GameLogic up 9.8%. Those compare Hy4 with its own earlier version, not rivals.

Finance and office. Tencent Financial Technology ran a crowdsourced test with 213 experts on 579 financial tasks. On a four-point scale Hy4 scored 3.04, against 2.97 for DeepSeek-V4.1-Flash and 2.68 each for Kimi K3.1 Pro and GLM-5.3. The lead over DeepSeek is small. Tencent also reports Hy4's hallucination rate in internal search tests fell from 15.76% to 5.45%, and the error rate from guessing from 23.1% to 5.6%. Partner Manus reports close to 97% completion on office document tasks, a vendor-reported figure.

How Hy4 ranks on Tencent's own benchmarks

Tencent reran rival models under consistent settings, so these are not scores published by each vendor and are not an independent leaderboard. Seven models were compared. Figures below are as reported by Apiyi from Tencent's chart.

Apiyi's reading: Hy4 wins six of twelve against Kimi K3.1 Pro and loses six. Its strengths cluster in code generation, 3D modeling, retrieval and research. Kimi leads on terminal work and automation.

The supplementary results are less flattering. On Hy-FrontierMath T5, Hy4 scored 13.0 against 97.5 for GPT-6.1 Sol. On Goal2Repo it scored 23.2 against 74.0 for Claude, and on Foundation Model Bench 28.1 against DeepSeek's 36.3. Tencent itself acknowledges shortcomings, which Apiyi says are listed in the release.

The Terminal Bench figure also needs care. One aggregator reported the jump as 26.3 to 44.4 from preview to final, while our earlier preview coverage cited a different terminal benchmark version. Check the benchmark name and version before comparing numbers across posts.

A hand placing a tile into a compartmented tray, picturing agent tasks completed step by step in software

Price and where to run it

Tencent's list price is RMB 6 per million input tokens, RMB 18 per million output tokens and RMB 0.3 per million cached tokens. Calls through Tencent Cloud TokenHub are 50% off through December 31, 2026. ProPakistani's coverage of the August preview listed the dollar price as $0.834 input and $2.501 output per million tokens. Coverage of the final release says the RMB prices are unchanged from the preview, so those dollar figures should still be close, but check the live price page before you budget.

The OpenRouter page reportedly shows a median throughput near 63 tokens per second, a median time to first token near 2.9 seconds and 99.87% availability over three days. The model is also reported live in CodeBuddy and WorkBuddy, and WorkBuddy's WeCom suite lets it read chats, email, documents, calendars and meetings. API compatibility, based on the preview docs, covers OpenAI Chat Completions, OpenAI Responses and Anthropic Messages formats.

Are the weights open?

Not yet for the final model. The coverage we read agrees that Tencent describes the weights as planned for open source, has partnered with vLLM and SGLang for support, and intends to publish a system card with safety evaluations alongside the weights. A smaller "Flash" version is reportedly planned. The August preview weights are on Hugging Face under Apache 2.0, per the coverage. We have not checked a download page for the final release, so do not assume a date.

This matters for the "open model" label. Until the weights ship, Hy4 is an API model with a promise attached. See our look at how Asian open models have taken a large share of OpenRouter tokens for why that promise is worth tracking.

What early users say

Apiyi collected a few day-one reactions from X, and it cautions that the sample is tiny. One user measured 60 to 70 tokens per second on WorkBuddy and called it "not fast". Another said real-world performance fell short of the benchmark scores. A third called results "so-so". These are individual impressions from launch day and not tests we reproduced.

Which model for which job

For the wider picture of open models, see our posts on Kimi K3's open weights and the GLM-5.3 intelligence index tie.

How to try it safely

1. Run your own 20 to 50 tasks, since vendor benchmarks do not cover your codebase.
2. Keep computer-use runs in a sandbox or virtual machine with no real credentials, as the model can click and type.
3. Log tool calls and cost per task, and compare with your current model.
4. Handle the reasoning_content field in your client and set output limits, because thinking is on by default.
5. Wait for the system card before sending sensitive data.

What is verified and what is not

• Reported by several outlets: release date, 770B size, 1M context, vision and computer use, price in RMB.
• Tencent's own data, not independent: every benchmark, the finance study, the hallucination rates and the 19.1% game gain.
• Not verified by us: the weights timeline, the demo details and OpenRouter statistics quoted from secondary write-ups, and Tencent's official page, which did not render for us.

Related reading

• Tencent Hy4 preview: 770B MoE with 1M context
• Tencent Hy3: 295B open-source agentic MoE
• Asia's open models and OpenRouter token share
• Kimi K3 open weights
• GLM-5.3 ties Kimi K3 on the intelligence index
• Tencent leases 100,000 AI chips from Oracle

Details are accurate as of November 4, 2026. Benchmarks and prices change quickly.