Tencent Hunyuan

Learning Objectives

• Understand Tencent Hunyuan's role in the Chinese AI ecosystem and global open model landscape
• Identify the Hy4 preview's mixture-of-experts architecture and multimodal capabilities
• Read a Hunyuan license correctly — the terms differ per checkpoint, and one of them excludes whole regions
• Evaluate when Hunyuan fits a deployment vs Western foundation models or other Chinese providers

What Is Tencent Hunyuan?

Tencent Hunyuan is the multimodal foundation model family from one of China's two AI superpowers (alongside Alibaba). Hunyuan powers AI across Tencent's massive consumer ecosystem — WeChat (1.3 billion+ monthly active users), QQ, and Tencent Cloud — supporting text generation, image creation, video understanding, and increasingly video generation.

The current flagship is the Hy4 preview, published on August 28, 2026: a mixture-of-experts (MoE) model with 770 billion total parameters and roughly 49 billion activated per token, with a context window past one million tokens. It is aimed at coding, office work, data analysis, game development and scientific research, and it ships with open weights on Hugging Face, ModelScope, GitCode and CNB, alongside a lower-precision FP8 variant. It reaches users through WorkBuddy, CodeBuddy, Yuanbao and ima, with API access via Tencent Cloud.

Tencent reports an internal blind evaluation — 163 experts across 203 engineering tasks — scoring Hy4 at 2.99 out of 4.00, against 2.92 for GLM-5.3 and 2.94 for Kimi K3. That is a vendor-run test on a vendor-authored benchmark, so treat it as a claim of rough parity with the open frontier rather than an independently measured win.

The prior flagship, Hy3 preview (April 23, 2026), was a 295 billion parameter MoE activating 21 billion per token and combining "fast and slow thinking" in one architecture. Elsewhere in the family, HunyuanImage 3.0-Instruct (January 2026) runs at 80 billion parameters and HunyuanVideo generates 16-second videos from text or image prompts.

Licensing — Read the Checkpoint, Not the Family

Hunyuan is the sharpest example in the catalog of why "open weights" and "open source" are not the same claim. Two Hunyuan checkpoints published on the same day carry different licenses, and one of them is unusable in whole regions.

Hy4 preview ships under plain Apache 2.0. Its LICENSE file is the standard Apache text with no addendum — no territorial carve-out, no revenue trigger, no monthly-user threshold, no acceptable-use rider. For a developer, that is as permissive as an open-weights release gets.

Most earlier Hunyuan releases ship under the Tencent Hunyuan Community License instead. That agreement defines its "Territory" as the world excluding the European Union, the United Kingdom and South Korea, states plainly that it does not apply in those regions, and forbids using the model, its derivatives, or its outputs outside the Territory. It also requires a separate written agreement from Tencent for any product whose monthly active users exceed 100 million, and it binds you to an acceptable-use policy.

The practical reading: the hosted API is globally reachable regardless of which license a checkpoint carries, because that is a service contract rather than a weights grant. The restriction bites when you download and self-host. If you are in an excluded region and the checkpoint you want is Community-licensed, the API path stays open and the self-hosting path does not.

Pricing & Access

Hunyuan is sold through Tencent Cloud APIs and integrated into Tencent's consumer products. Access for Western developers is more limited than for OpenAI / Anthropic / Google APIs due to regional availability and account requirements.

Open-source Hunyuan releases (HunyuanImage 3.0, HunyuanVideo) are available on Hugging Face and the Tencent-Hunyuan GitHub organization — meaningful for global developers wanting to deploy Hunyuan models without Tencent Cloud account dependencies.

Core Models

Hy4 preview (August 2026)

The current flagship, and a step change in size over Hy3. Mixture-of-experts (MoE) architecture:

• 770 billion total parameters — roughly two-and-a-half times Hy3's count
• 49 billion activated per token (efficient inference relative to model size)
• Context window past one million tokens
• Open weights on Hugging Face, ModelScope, GitCode and CNB, plus an FP8 variant for lower-precision serving
• Plain Apache 2.0 — see the licensing section above, because its siblings differ

Tencent positions it for coding, office work, data analysis, game development and scientific research, and reports it edging GLM-5.3 and Kimi K3 on an internal expert evaluation. It is a self-reported result; the honest summary is that Hy4 has joined the open frontier tier rather than that it leads it.

Hunyuan 3.0 / Hy3 preview (April 2026) — prior generation

The flagship until August 2026. Mixture-of-experts architecture combining:

• 295 billion total parameters
• 21 billion activated parameters per token
• "Fast and slow thinking" combined — adaptive reasoning depth based on query complexity
• Multimodal integration — audio + video + text in unified architecture

The MoE approach is broadly similar to DeepSeek's and Alibaba's — using sparsity to make very large parameter counts economical to inference.

HunyuanImage 3.0-Instruct (January 2026)

Image generation flagship. 80 billion total parameters, 13 billion activated per token. Native multimodal architecture — handles text-to-image, image-to-image, and image understanding in one model. Open-source release on Hugging Face makes it accessible to global developers.

HunyuanVideo

Video generation. 13-billion-parameter diffusion transformer specifically designed for temporal generation tasks. Capable of producing 16-second videos from:

• Text prompts (text-to-video)
• Image prompts (image-to-video, animating still images)

Open-source release on Hugging Face. Competitive with Western video generation models for many use cases at meaningfully lower deployment cost.

WeChat AI Agent

Tencent is concurrently developing an AI agent for WeChat designed to compete with cross-platform AI assistants. Given WeChat's 1.3 billion+ user base, this represents one of the largest single AI deployments globally.

Yuanbao + QClaw + WorkBuddy

Tencent's expanding AI product portfolio:

• Yuanbao — Tencent's consumer AI assistant
• QClaw — agentic AI capabilities (similar to OpenClaw class)
• WorkBuddy — workplace productivity AI

All built on Hunyuan foundation models.

NVIDIA TensorRT-LLM Optimization

Tencent works with NVIDIA to build high-performance inference engines for Hunyuan using TensorRT-LLM — meaningful for deployment efficiency, particularly given China's complex GPU supply situation.

Strengths

• Massive deployment scale: Hunyuan powers AI for 1.3 billion-plus WeChat users — distribution scale unmatched outside China
• Genuinely permissive flagship license: Hy4's plain Apache 2.0 carries no territorial, revenue, or user limits, which is rarer among frontier-scale open weights than the phrase "open" suggests
• MoE efficiency: 770 billion total against 49 billion activated per token puts frontier-scale capability within reach of practical inference budgets
• Long context: past one million tokens on the current flagship
• Multimodal native: text, image, video and audio in a unified architecture
• HunyuanVideo open weights: a 13 billion-parameter video generation model available to global developers
• HunyuanImage 3.0 open weights: 80 billion-parameter image generation
• Tencent Cloud integration: tight integration with Tencent's broader cloud platform
• NVIDIA partnership: TensorRT-LLM optimization for inference performance

Limitations & Considerations

• Regional availability: Tencent Cloud APIs primarily target Chinese and APAC markets; Western developer adoption smaller
• Geopolitical considerations: Chinese AI providers face different regulatory and procurement constraints in Western markets
• License terms vary per checkpoint, and one variant excludes whole regions: the Community License covering most non-Hy4 releases does not apply in the European Union, the United Kingdom or South Korea. Open the LICENSE file on the specific repository before you self-host
• English-language quality: Hunyuan is strong in Chinese; English performance generally trails frontier Western models
• Less Western developer ecosystem: smaller community of tutorials, integrations, and tooling outside China
• Frontier comparison rests on a vendor benchmark: Hy4's edge over GLM-5.3 and Kimi K3 comes from Tencent's own internal evaluation, with no independent replication yet
• Preview status: Hy4 is published as a preview rather than a settled release, so behaviour and endpoints may move
• Data handling concerns: Tencent Cloud data handling under Chinese law differs from US / EU privacy frameworks

Best Use Cases

When to choose alternatives:

• English-language workloads → a closed frontier flagship from OpenAI, Anthropic or Google for higher quality
• Western enterprise procurement → US-based providers may simplify compliance and procurement
• Open-weight Chinese alternatives → Qwen (Alibaba), DeepSeek, Kimi (Moonshot AI) also strong, with broader Western adoption
• Open-weight Western alternatives → Llama (Meta), Mistral for English-language workloads
• Deploying inside the EU, UK or South Korea → check the checkpoint's license first; a Community-licensed Hunyuan model is not available to you for self-hosting, while Qwen's Apache 2.0 releases and DeepSeek's MIT releases are

Key Takeaways

• Tencent Hunyuan is the multimodal foundation model family powering AI across Tencent's WeChat (1.3 billion-plus users), QQ, and Tencent Cloud
• The Hy4 preview (August 28, 2026) is the current flagship — a mixture-of-experts model with 770 billion total parameters, roughly 49 billion activated per token, and a context window past one million tokens. It supersedes the Hy3 preview, which was 295 billion total and 21 billion activated
• Hy4 ships under plain Apache 2.0 — no territorial carve-out, no revenue trigger, no user threshold. That is unusually permissive for a frontier-scale open-weights release
• The rest of the family does not share those terms. The Tencent Hunyuan Community License covering HunyuanOCR, HunyuanImage and HunyuanVideo excludes the European Union, the United Kingdom and South Korea from its Territory outright, and requires a separate Tencent agreement above 100 million monthly active users. Read the LICENSE file on the checkpoint you plan to run, not on its sibling
• Tencent's claim that Hy4 edges GLM-5.3 and Kimi K3 comes from its own internal expert evaluation — treat it as rough parity with the open frontier, not an independently measured win
• Best fit for Chinese-market AI applications, WeChat / QQ ecosystem integration, and self-hosted multimodal work; for English-language workloads or Western enterprise procurement, US providers and Western open-weight models often serve better