Tencent’s Hy3 preview Marks the First Major Model to Emerge From Its AI Rebuild

Tencent has introduced Hy3 preview, the first major model to emerge since the company rebuilt key parts of its pre-training and reinforcement-learning infrastructure earlier this year, making the release as much a sign of internal change as a routine model update. According to briefing materials, Hy3 preview reached public release in less than three months after that reset.

That timeline is notable in a market where new large language models often arrive in familiar form: bigger benchmarks, lower prices, broader claims. Hy3 preview comes with those elements too. Tencent describes it as a fast-and-slow-thinking fused MoE language model with 295 billion total parameters, 21 billion activated parameters, and support for up to 256K context. The company says the model has been designed for complex reasoning, instruction following, in-context learning, coding and agentic workloads.

Tencent’s chief AI scientist, Yao Shunyu, described HY3 Preview as the first step in rebuilding the Hunyuan model line, with the open-source release intended to bring in feedback from developers and users before the official HY3 version. Tencent said it is continuing to scale up pre-training and reinforcement learning in parallel, while using product co-design to improve the model’s real-world performance across scenarios.

What the Rebuild Actually Signals

But the more revealing part of the launch may be what it says about Tencent’s development cycle. Hy3 preview is being positioned as the first visible result of a broader rebuild around what the company calls more practical, utility-oriented models. The latest briefing emphasizes three ideas behind that shift: systematic capability, authentic evaluation and cost-effectiveness. Read another way, Tencent is trying to show that its model pipeline is no longer being optimized only for isolated performance gains, but for faster iteration into deployable systems.

Product Deployment as the Real Test

That product linkage appears early. Before launch, Tencent says Hy3 preview had already been deployed in a range of internal and consumer-facing products, including Yuanbao, CodeBuddy, WorkBuddy, ima, Tencent Docs and Peacekeeper Elite. In an industry where model quality is beginning to converge in headline terms, the question is increasingly not just who can train a capable model, but who can move updated models into real products quickly enough for the cycle of testing, feedback and iteration to matter.

Efficiency, Pricing, and Accessibility

Tencent is also pairing the launch with an efficiency story. Hy3 preview, according to the company, improves inference efficiency by 40%, and is being offered through Tencent Cloud’s TokenHub with pricing starting at RMB 1.2 per million input tokens and RMB 4 per million output tokens. That gives the release a second message beyond raw capability: Hy3 preview is being introduced as a model intended to be used broadly, not merely showcased.

Where This Fits in the Cycle

Hy3 preview may not be the most dramatic model launch in the current cycle. But it may be one of the more informative ones. For Tencent, it looks less like a standalone announcement than an early signal that a rebuilt AI pipeline is beginning to produce faster, more product-oriented results. Tencent said the launch will also be accompanied by a two-week period of free token access, giving users an early chance to test the first model released after the company’s latest AI rebuild.

Hy3 preview can be accessed at: https://openrouter.ai/tencent/hy3-preview:free (free access is available for a limited two-week period).