Tencent Launches Hy4: 770B-Parameter Open-Weight Model With Vision and Computer Use

Tencent released Hunyuan Hy4 on November 5, 2026 — a 770-billion-parameter mixture-of-experts model that marks the first time the Hy series ships with vision understanding and computer-use capabilities. It is the largest openly released Chinese AI model to date, and its combination of open weights, long-context reasoning, and multimodal action puts it in direct competition with GPT-4o and Gemini 2.0 Flash.

Tencent's Hunyuan line has been a commercial backbone for the company's enterprise AI products, but prior releases were text-only and kept weights proprietary. Hy4 changes both of those things. The company is making the weights available to the public within days of the launch announcement and is publishing a system card — a technical disclosure document describing training methods, safety evaluations, and known limitations — alongside the release.

What Hy4 Actually Is

The model uses a mixture-of-experts (MoE) architecture, a design where only a fraction of the model's parameters activate on any given input. Despite having 770 billion total parameters, Hy4 uses roughly 49 billion active parameters per token — keeping inference costs and latency manageable while preserving the breadth of knowledge packed into the full model.

Key technical specs per Tencent's announcement via Caixin Global:

• 770B total parameters / ~49B active per token
• 1 million token context window — roughly 750,000 words, or several novels
• Vision understanding: reads images, diagrams, charts, and screenshots
• Computer use: can operate software interfaces directly, clicking, typing, and navigating
• Hallucination rate: cut from 12.95% to 7.2% versus the prior generation
• Pricing: ¥6 per million input tokens / ¥18 per million output tokens (unchanged from the preview)

The 1M-token context window is significant. It means enterprises can feed entire codebases, legal contracts, or research archives into a single conversation without truncation.

Why the Hallucination Number Matters

A hallucination rate of 7.2% means that in roughly 1 in 14 responses, the model asserts something factually incorrect with apparent confidence. That is meaningfully better than the 12.95% rate of the prior Hy generation, but still above what most regulated industries — finance, healthcare, law — require before deploying a model in production. Tencent has not published the specific benchmark used to calculate these figures, which makes direct comparison to competing models difficult.

Open Weights: Who Gets What

Releasing open weights means that any organization can download the model's learned parameters and run Hy4 on their own hardware, without paying API fees or routing data through Tencent's servers. That matters for:

• Enterprises with data privacy requirements who cannot send proprietary information to an external API
• Researchers who want to study or fine-tune the model
• Regional AI builders in markets where Tencent's commercial API may face regulatory friction

The computer-use capability — the model's ability to control software directly — combined with open weights is a combination that has not appeared in any prior major open release. It raises obvious questions about misuse, which the forthcoming system card is expected to address.

What This Means for the AI Model Market

Tencent's move continues a trend that began with Meta's Llama releases: Chinese technology companies are open-sourcing frontier-scale models as a strategy to build developer ecosystems and pressure Western API providers on price. DeepSeek's V4 Flash, released at a fraction of comparable model costs in September, already forced several rounds of OpenAI and Anthropic price reductions.

Hy4's pricing — roughly $0.83 input / $2.50 output per million tokens at current exchange rates — sits well below GPT-4o and Claude Opus pricing for equivalent capability tiers, though direct performance benchmarks have not yet been independently verified.

For businesses currently paying premium rates for multimodal AI with long context windows, Hy4's commercial API provides immediate leverage in vendor negotiations, regardless of whether they switch.

What to Watch

Tencent says weights will be available "within days." The key question is what the system card discloses about training data sourcing, safety red-teaming results, and the specific compute required to run the full model. For enterprise adoption outside China, the computer-use capability will likely face the most scrutiny — giving a model the ability to click and type on live systems is a significant trust boundary that most IT security teams will want independently evaluated before deployment.