Tencent ships Hy3 as an Apache 2.0 agent model

Why it matters

Tencent is using Hy3 to turn internal product telemetry into an open model distribution play, with Apache 2.0 licensing and OpenRouter access aimed at agent developers who may never touch Tencent Cloud directly.

Tencent's Tencent Hy (@TencentHunyuan) team released Hy3 on July 6th, saying in a post on X that the 295 billion-parameter mixture-of-experts model is available under Apache 2.0, with two weeks of free API access through OpenRouter.

https://x.com/TencentHunyuan/status/2074148098876768478

The release turns Hy3 from an April preview into a fuller open-weight push for agent developers. Tencent's official announcement says Hy3 is already used across WorkBuddy/CodeBuddy, Yuanbao, Marvis, ima and other Tencent products, and that the API is available on Tencent Cloud TokenHub. The model weights are listed on Hugging Face, with a GitHub repository and a research page carrying the technical description.

The technical pitch is efficiency at scale. Tencent describes Hy3 as a 295B-parameter MoE model with 21B active parameters, 3.8B MTP layer parameters, 80 layers excluding the MTP layer, 192 experts with top-8 activation, BF16 support and a 256K context length. The Hugging Face model card lists the license as Apache 2.0 and gives serving instructions for vLLM and SGLang, including OpenAI-compatible API calls after deployment.

That licensing change matters because Tencent is trying to make Hy3 easier to adopt outside its own cloud and product stack. The July 6th announcement says Hy3 and Hy3-FP8 weights are being open sourced across Hugging Face, ModelScope, GitCode and CNB. Tencent also says global third-party developer platforms including OpenRouter, Hermes, Kilo, Cline, OpenClaw, OpenCode and Cherry Studio will integrate Hy3 progressively. OpenRouter already lists Tencent: Hy3 (free), with a model-weights link back to Hugging Face.

Tencent is making a broad performance claim, and the wording should be read as company-supplied. The Hy3 model card says the model outperforms similar-size models and rivals flagship open-source models with two to five times the parameters. It also says Tencent ran a blind test with 270 experts and collected 312 valid comparisons, with Hy3 scoring 2.67 out of 4 versus GLM-5.1 at 2.51. Those comparisons are useful as a window into Tencent's target market - coding, CI/CD, frontend development, data and storage work - but they are not the same as an independently replicated benchmark suite.

The product feedback loop is the sharper part of Tencent's case. Tencent says the team gathered feedback from more than 50 product teams after the late-April preview, fixed task-execution and interaction issues, and scaled post-training with higher-quality data. In its corporate announcement, Tencent says average daily token consumption for Hy3 preview increased twentyfold after the preview launch. It also says WorkBuddy users who actively select Hy3 preview grew sixfold. Those are internal Tencent metrics, and Tencent does not publish the underlying token base or user denominator, so they show direction rather than market scale.

Yao Shunyu, Tencent's chief AI scientist, framed the April preview as a way to collect feedback before the official launch. In Tencent's April 24th preview announcement, Yao said the release would help Tencent optimize Hy3's performance and real-world applicability through product co-design. That is exactly the strategy visible in the July release: Tencent is treating its own apps as both distribution and evaluation infrastructure.

Hy3 also shows how Chinese AI labs are competing on deployable open models rather than pure parameter-count theater. Tencent says Hy3 preview was integrated into Yuanbao, ima, CodeBuddy, WorkBuddy, QQ, QQ Browser, Tencent Docs and Tencent LearnShare. The July release adds more concrete internal use cases: Yuanbao's agent function can generate files such as PowerPoint, Word, Excel, PDF and HTML; Marvis uses Hy3 for file editing, file management and computer diagnostics; Weixin Official Accounts customer-service assistants use it for context-aware responses; and Path of Exile: Advent on WeGame has connected an AI assistant to Hy3.

Tencent's reliability claims are aimed at developers building agents, where a strong chat score is less important than stable tool calls and predictable output structure. The model card says Tencent improved tool-call stability and error recovery, reduced invalid calls that can trigger infinite loops, and kept SWE-Bench Verified accuracy variance within 4% across scaffolds including CodeBuddy, Cline and KiloCode. It also says internal real-world evaluations showed Hy3's hallucination rate dropping from 12.5% to 5.4%, commonsense error rates falling from 25.4% to 12.7%, and multi-turn issue rates declining from 17.4% to 7.9%. Tencent does not provide enough public detail in the model card to independently audit those internal evaluations.

The immediate constraint is deployment. The model card says Hy3 has 295B total parameters and recommends H20-3e or other large-memory GPUs for serving on 8 GPUs. That limits who can self-host the full model, even with only 21B active parameters per token. Tencent's OpenRouter free period and third-party platform push address that gap by putting the model behind APIs while it tries to seed developer usage.

For Tencent, Hy3 is also an answer to a distribution problem. The company has consumer surfaces, enterprise cloud channels and internal productivity products, but the global developer mindshare around agent frameworks has been harder to win. Apache 2.0 weights, Hugging Face availability and OpenRouter access give Tencent Hy a cleaner route into the agent tooling market than a China-only cloud endpoint or a restrictive research release.

The open question is whether outside developers will see the same stability Tencent reports from its own product estate. Hy3 has the ingredients developers ask for: long context, open weights, a permissive license, MoE efficiency, FP8 availability and an API path. The test starts when builders run it inside their own agents, with their own tools, prompts, evals and failure modes.