Tencent Open-Sources Hy3-Preview: Frontier MoE Built in 3 Months Under Ex-OpenAI Researcher
Tencent released Hy3-preview, a 295B parameter MoE language model (21B active per token) with 256K context, as open weights on Hugging Face. Built in under three months under former OpenAI researcher Yao Shunyu, the model targets agentic and coding use cases and is available for free inference on OpenRouter. It joins GLM-5.1 and Qwen3.6 in the race to build frontier open-source models outside of Nvidia's supply chain.
Original sourceTencent released Hy3-preview on April 23, 2026 — its first publicly available frontier-class language model — as open weights on Hugging Face. The release marks the company's formal entry into the competitive open-source LLM race, alongside Alibaba's Qwen, Zhipu's GLM, and Xiaomi's MiMo series.
The architecture is a 295B parameter Mixture-of-Experts model with 192 routed experts, keeping only 21B parameters active per token during inference. This efficiency profile makes Hy3 accessible for organizations with standard GPU clusters rather than requiring specialized infrastructure. The model supports a 256K context window and was tuned specifically for agentic task completion and software engineering benchmarks.
The release comes six months after Yao Shunyu — a notable AI researcher who spent several years at OpenAI — joined Tencent to lead its frontier model efforts. The team claims Hy3-preview went from project kickoff to public release in under 90 days, an aggressive timeline for a model at this scale. Tencent has not yet published a full technical report, so training data composition and methodology remain partially opaque.
Free inference is available at launch on OpenRouter's free tier, with rate limits. The model is also accessible via Hugging Face's Inference API. For the broader open-source ecosystem, Hy3-preview adds another credible frontier option — particularly relevant for developers who need strong multilingual capabilities and Chinese-language performance, where Tencent's training data advantages are likely significant.
The release accelerates what is now a four-way open-source frontier race in China: Alibaba, Zhipu, Xiaomi, and Tencent are all shipping competitive models within weeks of each other. Combined with the fact that GLM-5.1 was trained on 100,000 Huawei chips with zero Nvidia hardware, these releases collectively challenge the assumption that frontier AI requires US-controlled semiconductor supply chains.
Panel Takes
The Builder
Developer Perspective
“Free on OpenRouter is the headline for practitioners — I can benchmark Hy3 against my current stack today without any commitment. The 256K context and agent-tuning make it worth serious evaluation for long-horizon task automation.”
The Skeptic
Reality Check
“No technical report means we're taking Tencent's benchmark claims on faith. The 'three months from zero to frontier' timeline raises eyebrows — either the methodology was very efficient or the benchmarks are being measured selectively. Wait for third-party evals before making infrastructure decisions.”
The Futurist
Big Picture
“Four competitive Chinese labs shipping frontier open-source models simultaneously is a geopolitical inflection point. The narrative that 'AI leadership requires US chips and US labs' is being visibly challenged — and open weights mean the capability diffuses globally regardless of export controls.”