Tencent’s Hy3 LLM Goes Global with Free Access Until 2026

Tencent officially launched its Hy3 large language model on July 6, 2026. The company wasted no time. On August 5, 2026, Tencent announced Hy3’s global availability across its products and cloud platforms.

Hy3 is no lightweight. It packs 295 billion total parameters and activates 21 billion at a time using Mixture-of-Experts architecture. It supports a context length up to 256,000 tokens — that’s a massive window for complex tasks.

Tencent claims Hy3 matches or outperforms flagship models with two to five times as many parameters. It excels in reasoning, instruction understanding, in-context learning, code generation, and agent tasks. That’s quite a feat for a model smaller than some competitors.

Users can access Hy3 free of charge on WorkBuddy until August 31, 2026, Pacific Time. The API is also available on OpenRouter, priced at $0.1288 per million input tokens and $0.5336 per million output tokens.

Hy3’s weights are open source under the Apache 2.0 license. Tencent published them on Hugging Face and ModelScope, inviting developers to build on its foundation. This move signals Tencent’s intent to compete on transparency and accessibility.

Usage stats tell a story. Daily token consumption of Hy3 surged twenty-fold since its preview launch on April 23, 2026. Within a week of official release, API calls exceeded the previous generation model by 68 times. Hy3 even topped OpenRouter’s global LLM usage leaderboard in that short span.

Tencent integrated Hy3 deeply. It powers Yuanbao assistant, the ima knowledge-base tool, Marvis OS-level agent, and customer-service bots on Weixin Official Accounts. This rollout shows Tencent’s push to embed AI across its ecosystem, not just offer a standalone product.

Poshu Yeung, Senior Vice President of Tencent Cloud International, called Hy3 a “significant step” toward moving organizations beyond AI experimentation. He emphasized practical, trusted, and scalable intelligence as the goal.

Context in a Crowded Field

Tencent’s Hy3 arrives amid stiff competition. Alibaba’s Qwen team unveiled Qwen3.8-Max on August 3, 2026. It’s a 2.4-trillion-parameter multimodal Mixture-of-Experts model. Qwen3.8-Max leads several benchmarks, including OSWorld-Verified at 86.1 and PaperBench at 93.0, surpassing GPT-5.6 Sol Max and Anthropic’s Fable 5.

Alibaba plans to release Qwen3.8-Max’s open weights next week alongside Qwen3.8-27B. Pricing on QwenCloud runs $2 per million input/output tokens, totaling $8 per million tokens. Its inference costs undercut competitors like Claude Opus 5 and GPT-5.6 Sol Max by a wide margin.

Meanwhile, Moonshot AI launched Kimi K3 on July 27, 2026. It’s a free, customizable AI model supporting multiple languages and image reading. Kimi K3 ranks third on Artificial Analysis’s AI systems index. It aims to help governments bypass costly licenses on U.S.-controlled cloud platforms.

Tencent’s Hy3, Alibaba’s Qwen3.8-Max, and Moonshot’s Kimi K3 target different niches but share a common goal: open access and practical AI adoption. Tencent’s hybrid approach—open weights, free use on WorkBuddy, and broad product integration—positions Hy3 as a flexible, enterprise-ready option.

This global push by Tencent highlights China’s accelerating AI ambitions. With Hy3, Tencent stakes a claim in the international AI arena. Whether it can keep pace with larger, more expensive models remains to be seen. But for now, Hy3’s mix of performance, openness, and broad availability makes it a major player worth watching.

Stay connected via Google News

Follow us for the latest AI News