Most big-tech AI launches are demos dressed up as products. Tencent spent the early part of 2026 doing the opposite — quietly rebuilding its model from the training stack up, then handing the result to the public for free. The bet is simple: usefulness beats spectacle, and an open model that lives inside every app is harder to leave than one you only visit on a website.
A model rebuilt for work, not leaderboards
On 23 April 2026, Tencent released and open-sourced Hy3 preview, the flagship of its Hunyuan (混元) series. It is a Mixture-of-Experts model that blends "fast" and "slow" thinking, with 295 billion total parameters, 21 billion activated, and a context window of 256K tokens.
Chief AI Scientist Yao Shunyu (姚顺雨) said the release was a way to "gather feedback from users and the broader community" ahead of the formal launch, and that Tencent is "continuously expanding the scale of our pre-training and reinforcement learning efforts." That is not marketing filler: since February 2026 the company says it rebuilt its pre-training and RL infrastructure around three principles — well-rounded capability across reasoning, long context and tool use; evaluation that looks beyond standard benchmarks; and tight co-design between model and product so cost stays low.
The explicit goal was productivity, not a higher score on a public chart.
Inside the apps: Yuanbao and the work stack
The more interesting move is where Hy3 preview landed. Tencent did not park it on a showcase page — it pushed the model into the products hundreds of millions already use:
- Yuanbao (元宝), the consumer AI assistant, got sharper intent understanding and text quality from deep co-design with the model.
- CodeBuddy and WorkBuddy — Tencent's coding and agent tools — saw first-token latency drop 54%, end-to-end response time fall 47%, and a success rate above 99.99%. In real user environments the model reportedly sustained agent workflows of up to 495 steps, covering document processing, data analysis, knowledge retrieval and Model Context Protocol (MCP) tool orchestration.
- Tencent Docs (腾讯文档) AI PPT feature improved generation success by 20% versus the previous Hy2 model.
- The model also reached ima, QQ, QQ Browser and Tencent LearnShare (腾讯乐享), with more products in the pipeline.
The pattern is clear: Tencent is treating its foundation model less like a product and more like an operating layer.
Cheap to run, free to fork
Efficiency is the part competitors will feel. Tencent says Hy3 preview delivers a 40% improvement in inference efficiency at comparable cost, thanks to co-optimization between model architecture and inference stack.
Pricing on Tencent Cloud's TokenHub platform:
- Input from about US$0.18 per million tokens.
- Cached input from about US$0.06.
- Output from about US$0.59.
- A personal plan from roughly US$4.10 per month for use inside agent frameworks such as OpenClaw.
And it is genuinely open. Hy3 preview is published on GitHub, Hugging Face, ModelScope and GitCode, supports mainstream inference frameworks like vLLM and SGLang, and the API was listed on OpenRouter with a two-week free window. Separately, Tencent open-sourced Hy-MT2-30B-A3B, a translation model, on 21 May 2026, covering 33 languages including minority-language Chinese translation.
Why give it away?
Open-sourcing a flagship is a strategic choice, not charity. In a market where open-weight Chinese models have already reset user expectations, giving developers the weights pulls them into Tencent Cloud's ecosystem — they fine-tune, deploy and pay for tokens on Tencent's infrastructure. Embedding the same model in Yuanbao and the office suite turns everyday usage into a distribution channel no standalone chatbot can match. The risk Tencent is managing is relevance: a model people only visit is easy to replace; one wired into where work happens is not.
Honest limitations
Primary sources are Tencent's official newsroom and the Hunyuan site. The performance figures — 54% lower latency, 99.99% success rate, 495-step workflows, 20% better PPT generation — are Tencent's own benchmarks and product-test claims, not independently audited results, and "real user environments" are described by Tencent rather than measured by a third party. The 295B / 21B parameter and 256K context numbers are vendor specifications. USD prices are converted from Tencent's RMB list (¥1.2 / ¥0.4 / ¥4 per million tokens; a ¥28 monthly plan), so they move with the exchange rate. "Open source" here means model weights released under Tencent's license terms, which developers should read before commercial use — it is not automatically a fully permissive license. Availability of Yuanbao (元宝) and some features varies by region.
What readers can do now
- Pull the weights and benchmark on your own tasks. Hy3 preview is on Hugging Face and ModelScope — test it against your real workloads instead of trusting the vendor leaderboard.
- Try the agent claims cheaply. Use the TokenHub free tier or the OpenRouter window to check whether the 495-step workflow claim holds for your pipelines.
- For Chinese↔multilingual needs, A/B test Hy-MT2 against commercial translation APIs before committing, since 33-language coverage is exactly where open models often still slip.