Cinevva

English

简体中文
한국어
हिन्दी

English

简体中文
한국어
हिन्दी

Appearance

Tencent open-sources Hunyuan Hy3, and it's already working inside Path of Exile ​

Tencent released Hunyuan Hy3 on July 6, and the headline isn't the benchmarks. It's the license. The full model, a 295-billion-parameter Mixture-of-Experts with 21 billion active per token, shipped under Apache 2.0 with no field-of-use or geographic restrictions. The April preview came with a restrictive license, and Tencent reversed that for the final release. You can now take a frontier-class Chinese reasoning model, put it in a commercial product, and owe nobody anything.

What Tencent actually shipped ​

Hy3 is a hybrid fast-and-slow-thinking model. It answers simple queries directly and switches into extended reasoning when the problem needs it, all in one system rather than two model variants. Context runs to 256K tokens. The weights are up on Hugging Face and ModelScope, with a rollout underway across OpenRouter and a stack of coding tools like Cline and Cherry Studio.

The pricing is the other half of the story. Through Tencent Cloud's TokenHub API, input tokens cost 1 yuan per million, about 15 US cents, with output at 4 yuan and cached input at a quarter of a yuan. Tencent says the model hits a 90% task-completion rate on agent workloads across its internal applications, and it has already been wired into Yuanbao, WeChat, ima, and the WorkBuddy coding assistant. Independent testing puts it ahead of other open models on agentic search and tool use, though it trails competitors like GLM-5.2 on some coding benchmarks, so it's a strong release rather than a clean sweep.

The Path of Exile deployment ​

Here's the part that matters for games. According to Tencent's announcement, the AI player assistant in Path of Exile: Advent on WeGame now runs on Hy3. The upgrade makes the assistant respond to player requests more accurately while cutting hallucinations, which is exactly the failure mode that has kept in-game AI helpers feeling like a gimmick. Path of Exile is a game whose build system is so deep that players keep wikis open on a second monitor. An assistant that actually understands the question is a real feature there, not a demo.

That's a concrete answer to a question the industry has been circling for two years: what does a language model actually do inside a shipped game? Not generate the game. Sit inside it, know its systems, and help the player. At 1 yuan per million input tokens, the economics of that feature just changed. An assistant that would have been a meaningful line item on a live-service budget a year ago is now cheap enough to be table stakes.

What it means for creators ​

We covered Tencent's Hunyuan 3D engine when it went global last November, and the pattern is the same one repeating a level up. First the asset pipeline commoditized. Now the intelligence layer is doing it too. Apache 2.0 weights at this quality mean any studio, any tool, any solo developer can run a frontier reasoning model on their own terms, fine-tune it on their game's systems, and ship an NPC or an assistant without a per-seat license or an API dependency they don't control.

For Cinevva this is the tailwind we build against. Every drop in the cost of intelligence makes AI features inside games, smart NPCs, in-game helpers, dynamic dialogue, less of an AAA luxury and more of a default. The gap between what a big studio can afford to put in a game and what you can is closing from both ends. Hy3 just closed it a bit more.

References ​