Tencent's Hy4 Goes Live: 770B Parameters, Vision, and Computer Use Join the Open-Weight Race

What Tencent Announced on November 4

On November 4, 2026, Tencent switched on the general-availability build of Hunyuan Hy4, the new flagship of its large language model family and the company's most aggressive entry yet in China's open-weight AI race against DeepSeek, Z.ai, and Alibaba's Qwen team.

Hy4 is a 770-billion-parameter mixture-of-experts system — a 2.6-fold jump from the 295 billion parameters of its predecessor, Hy3. Ten weeks ago, the August preview of this model was text-only and built mostly for coding and office work. Today's release can see, reason over images, and operate software on its own.

The scale of Tencent's commitment is visible in its spending: the company raised capital expenditure by 176% year-on-year to 52.8 billion yuan (~US$7.9 billion) in Q2 to fund its AI push, and it has already signaled that the next-generation Hy5 will scale further.

Vision and Computer Use, for the First Time

The headline upgrade is vision. Hy4 can process images, analyze complex charts, and perform visual mathematical reasoning — combining perception, reasoning, and tool invocation for tasks like fine-grained image recognition. It can also reason across multiple images and interpret documents.

The second new capability is Computer Use: Hy4 can execute complex tasks by combining command-line and graphical interfaces. That places it in the same small category of models — alongside Anthropic's and OpenAI's computer-use agents — that can read a screen and operate it, rather than just answer questions about it.

For a model family that couldn't see an image at all in August, the jump is fast even by 2026 standards, and it mirrors the industry-wide shift from chatbots toward AI agents that act inside real software. If you're evaluating agentic tools, our guide to the best AI agents of 2026 covers the landscape this launch lands in.

The Benchmarks: Wins, Losses, and a Halved Hallucination Rate

Tencent's published internal evaluations show the official Hy4 significantly outperforming the August preview across 12 tasks spanning agentic capabilities, coding, and workplace automation — including a 19.1% gain on its GameDev benchmark and 9.8% on GameLogic.

Against rivals, Tencent's technical blog claims Hy4 outperformed Z.ai's GLM-5.3 and DeepSeek's V4.1-Flash-1017 across all tested tasks, while trading wins with Moonshot's Kimi K3.1 Pro. On an internal finance evaluation covering 579 real-world tasks — run with 213 professionals scoring the results — Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash at 2.97, with Kimi K3.1 Pro and GLM-5.3 at 2.68.

Perhaps the most practical number is hallucinations: in Tencent's internal search evaluations, the hallucination rate fell from 15.76% in the August preview to 5.45% in this release. Reliability, not raw benchmark scores, is what separates usable AI tools from demos.

Tencent was also candid about where Hy4 still falls short: the company acknowledged it trails the leading closed-source systems on research-level mathematics, long-horizon coding, model self-improvement, and advanced visual question answering, pledging to "continue to steadily scale pre-training and post-training to narrow this gap."

The Catch: Where Are the Weights?

Here's the twist. Tencent is marketing Hy4 as "one of the strongest open-weight models available today" — but at launch, it shipped API-only, with the model weights and system card listed as "planned for a later release."

That's a notable reversal of its usual playbook. The August preview was released under an Apache 2.0 license with downloadable weights on Hugging Face — the same open strategy DeepSeek, Qwen, and Kimi have used all year to win developer mindshare. Pulling the weights back, even temporarily, breaks that pattern at exactly the moment rivals keep shipping open checkpoints. Tencent says the open-source release is coming soon, alongside a lighter Hy4 Flash focused on inference speed and cost.

On safety, the company says training kept only defensive cybersecurity data (proof-of-concept generation and patching), filtered out exploit development and penetration testing, and that red-team audits by its Xuanwu and Keen Labs found no elevated offensive risk. Its biosecurity capability scored 30.4/100, below the high-risk threshold.

Pricing and Availability

Hy4 is available now through Tencent Cloud and OpenRouter, and integrated into Tencent's own products: TokenHub, WorkBuddy, CodeBuddy, Yuanbao, ima, and the WeCom intelligent assistant.

• List API pricing: 6 yuan per million input tokens, 18 yuan per million output tokens, 0.3 yuan per million cache-hit tokens (about $0.834/$2.501 per million on the international list)
• TokenHub discount: 50% off Hy4 from November 4 through December 31, 2026
• Free usage of WorkBuddy and CodeBuddy from November 4 to 15

Even at list price, that's a small fraction of what frontier closed models charge — continuing the price-compression trend that has defined the open-weight race all year, and the same dynamic behind the ongoing AI inference price war.

What It Means for the AI Tools You Use

Three practical takeaways for anyone choosing AI tools right now:

• Vision-plus-computer-use is going mainstream. What was a premium frontier-lab capability six months ago is now table stakes for flagship models. Expect coding agents and assistants built on Hy4 (like Tencent's CodeBuddy) to gain screen-reading and app-operating features quickly — compare this against established options like Claude Code and OpenAI Codex.
• "Open-weight" now has an asterisk. A model advertised as open-weight but shipped API-only is a marketing claim, not a guarantee. If self-hosting matters to you — on Hugging Face, your own GPUs, or a private cloud — verify that weights are actually downloadable before committing to a model.
• Hallucination rates are becoming a headline metric. Tencent leading its own announcement with a 15.76% → 5.45% cut signals that buyers are increasingly asking "how often is it wrong" rather than "what's its benchmark score." That's healthy.

The open-weight race has been the single biggest driver of falling AI costs and rising capability in 2026. Tencent's Hy4 — huge, cheap, multimodal, and agent-capable — keeps that pressure on, even with the weights question unresolved.

Frequently Asked Questions