Tencent officially launched Hunyuan 3, referred to internally as Hy3, on July 6 – marking the first major model release since the company recruited Yao Shunyu, a leading AI scientist who had been a research leader at OpenAI, as its Chief AI Scientist in December 2025. The model’s performance represents a substantial acceleration from where Tencent’s in-house AI capabilities stood before Yao’s arrival and his subsequent restructuring of the company’s AI research infrastructure. Hy3 uses a Mixture-of-Experts architecture with 295 billion total parameters and 21 billion active parameters per inference, a 256K context window, and has been fully integrated into Tencent’s portfolio of consumer and enterprise products including the Yuanbao consumer AI assistant, coding tool CodeBuddy, productivity platform WorkBuddy, and WeChat-connected applications. Tencent open-sourced the model under the Apache 2.0 license and simultaneously cut its API input price to 1 yuan per million tokens. NEWSCENTRAL reads the combination of benchmark performance and pricing as the clearest signal yet that Tencent has decided to compete for AI developer market share through cost positioning rather than through closed-model premium pricing.
The benchmark results that accompanied Hy3’s launch place it in a tier of model performance that would have been inconceivable for any Chinese lab’s internally developed model a year ago. On the BrowseComp search agent benchmark, Hy3 scored 84.2, matching GPT-5.5. On the FrontierScience-Olympiad scientific reasoning benchmark, it outperformed GPT-5.5 outright. On the GPQA Diamond reasoning benchmark, its score of 90.4 approached GPT-5.5’s 93.6. On SWE-Bench Pro, a coding agent benchmark, it scored 57.9. In Tencent’s own in-context learning benchmark CL-bench, it scored 23.8 – ranking first among Chinese models and second only to Anthropic’s Claude Opus 4.8 at 24.8. The breadth of the benchmark coverage reflects a deliberate attempt to demonstrate frontier capability across multiple task categories simultaneously rather than optimizing for a single high-visibility metric.
The organizational change underlying Hy3’s development is at least as significant as the technical results. Yao Shunyu rebuilt Tencent’s AI pre-training and reinforcement learning infrastructure from late January through to the April preview release of Hy3, then continued to improve the model through enhanced post-training data quality and expanded reinforcement learning compute over the following 74 days before the July 6 official launch. The daily token consumption of Hy3 grew 20-fold between the April preview and the July official launch, reflecting rapid adoption across Tencent’s internal products and through its external API. Tencent shares rose 4.82% on the day of the official launch. Freddy Miller, Senior Analyst at NEWSCENTRAL, argues that the speed of Hy3’s development – from infrastructure rebuild to competitive frontier model in approximately six months – provides important evidence for the broader debate about how quickly Chinese AI labs can close the capability gap with American frontier developers when they hire the right talent and provide adequate compute resources. The answer, in Tencent’s case, appears to be: considerably faster than most external observers had predicted.
The commercial positioning of Hy3 at 1 yuan per million input tokens places it among the cheapest frontier-competitive models available globally. For context, Anthropic’s Claude Sonnet 5 is priced at $3 per million input tokens at standard rates, and GPT-5.5 carries comparable pricing. At roughly 14 cents per million input tokens at current exchange rates, Hy3 represents a 20-to-1 cost advantage over the leading American frontier models for developers and enterprises that can use it. The open-source release under Apache 2.0 adds a further dimension: organizations that prefer to self-host rather than use the API can download and run the model without any per-token cost at all.
NEWSCENTRAL assesses the Hunyuan 3 launch as marking a transition point in the Chinese AI competitive landscape that the global enterprise market has not yet fully priced in. Six months ago, no Chinese model had demonstrated credible frontier-competitive performance across multiple agentic task categories simultaneously. Hy3 now has, and it has done so while being open-sourced and priced at a fraction of American alternatives. The enterprise procurement implications of that combination will take several quarters to fully materialize – but the direction is now visible.
Tencent’s decision to fully integrate Hy3 into WeChat – the super-app with over a billion monthly active users in China – before the model’s official public launch, combined with the simultaneous deployment across WorkBuddy and Yuanbao, gives the model an immediate and massive production deployment base. That deployment at scale is both a commercial advantage and a technical feedback mechanism: usage patterns across Tencent’s real-world business scenarios generate the behavioral data that drives continued model improvement, creating a virtuous cycle between deployment and capability that closed, API-only models cannot replicate. The OpenRouter developer platform recorded Tencent’s model family at 8.7% share of total token calls in June 2026, reflecting growing enterprise interest in the model before the official Hy3 launch. NEWS CENTRAL considers that pre-launch adoption figure the strongest leading indicator of how quickly the open model ecosystem will shift further toward Chinese alternatives as Hy3’s performance and pricing become more widely known.