Tencent open-sources Hy4 Preview, a 770-billion-parameter model

Tencent open-sourced its next-generation large language model, 'Hy4 Preview,' on August 28. Built on a mixture-of-experts (MoE) architecture, it holds 770 billion parameters in total, of which only 49 billion are activated for any given token. Its context window exceeds one million tokens, enough to hold an entire codebase or several long documents without losing track of earlier context. That marks a clear step up from earlier releases in Tencent's Hunyuan family, both in overall model size and in how much context it can handle at once.

Built for agents and productivity work, not just chat

Tencent says Hy4 Preview was designed not as a general chatbot but for coding agents, complex tool-use workflows, and productivity tasks that require planning, context continuity and multi-step execution. The company says it worked with domain experts in software engineering, game development, finance and security from the design stage onward, training the model on real work-oriented data. It is tightly integrated with Tencent's own tools — CodeBuddy for coding, WorkBuddy for office collaboration — with the stated goal of positioning it as a B2B productivity solution engineers and analysts can put to work immediately.

Edging out GLM-5.3 and Kimi K3 in a blind evaluation

Tencent ran an internal blind evaluation involving 163 experts across 203 real-world engineering tasks. Hy4 Preview scored an average of 2.99 out of 4.00, slightly ahead of Zhipu AI's GLM-5.3 (2.92) and Moonshot AI's Kimi K3 (2.94). Coding and agentic performance improved substantially over the previous version, putting the model at the top of open models — though Tencent itself acknowledges it still trails the most advanced closed models from OpenAI and Anthropic.

Tencent points out that China's AI market is seeing intensifying competition between large tech companies and startups, both on model performance and on getting products to market. Rather than waiting for a fully polished release, the company says it chose to ship a preview first to gather real market feedback and iterate quickly from there. The release lands in a stretch where Zhipu AI (with its GLM lineup) and Moonshot AI (with its Kimi lineup) have both been rolling out new models in quick succession — Tencent is positioning Hy4 Preview as a work-focused alternative tied to its own product ecosystem.

• 770 billion total parameters, 49 billion active (MoE architecture)
• context window exceeding one million tokens, enough for a full codebase or bulky documents
• available via WorkBuddy, CodeBuddy (Chinese and international versions), Yuanbao, ima, the Tencent Cloud TokenHub API, and OpenRouter
• free for two weeks on WorkBuddy and CodeBuddy at launch
• API pricing: $0.834 per million input tokens, $2.501 per million output tokens
• weights published simultaneously on Hugging Face, GitHub and ModelScope

Acknowledged limits and a fast-iteration strategy

Tencent acknowledges that, as a preview ahead of a final release, the model still has limitations: thinking time that stretches out on complex tasks, and a tendency toward excessive self-verification. The company also says the model helped optimize parts of its own training and inference systems, lifting end-to-end throughput by 31.8% against a baseline. The stated strategy is to ship a preview quickly to gather real market feedback rather than wait for a polished final version — while the free trial period on CodeBuddy and WorkBuddy doubles as a way to lock early users into Tencent's own ecosystem.

What this changes for companies

For engineering teams outside China, Hy4 Preview adds one more option that can be reached without dedicated infrastructure, through OpenRouter or the Tencent Cloud API, and its open weights make it a candidate for teams weighing self-hosted deployment or their own fine-tuning. At under a dollar per million input tokens, its API pricing could also matter for teams processing large volumes of code or documents. Still, a lead of only a few decimal points over GLM-5.3 and Kimi K3 shows how tight the race among Chinese open models has become: before any production use, running the model against real internal tasks remains the only reliable way to judge whether it fits, especially given data-governance considerations that come with any model choice.