Tencent

Artificial Intelligence · Open-Source LLM · Developer Tools

Launch

Tencent open-sources 770B-parameter Hy4, claims it outperforms Codex

Free access for two weeks and rock-bottom API pricing aim to peel developers away from closed frontier models.

• Tencent's Hunyuan team open-sourced Hy4 preview on Aug. 28, a 770-billion-parameter Mixture-of-Experts model with 49B active parameters per token and a context window over 1 million tokens.
• Tencent built Hy4 with training data drawn from its own software engineers, game developers, finance analysts and security experts, aiming it at real workplace productivity tasks rather than pure chat.
• Tencent says Hy4 can direct multiple OpenAI Codex sessions as sub-agents and outperform Codex used alone across benchmarks it selected, a claim that has not been independently verified.
• In Tencent's own blind evaluation of 163 experts across 203 engineering tasks, Hy4 scored 2.99 out of 4, narrowly ahead of Zhipu's GLM-5.3 (2.92) and Moonshot's Kimi K3 (2.94).
• API access is priced at $0.834 per million input tokens and $2.501 per million output tokens, with two weeks of free access inside Tencent's WorkBuddy and CodeBuddy apps.
• Hy4 lands less than two months after Hy3 (295B total, 21B active parameters, launched July 6), a much faster upgrade cadence for a company widely seen as trailing rivals in China's AI race.
• Tencent's productivity comparison was scored by its own employees on tasks it selected, so the claims of beating Codex, GLM and Kimi remain company-reported rather than independently confirmed.
• Giving away a frontier-scale model with cheap API pricing pressures closed labs like OpenAI and Anthropic and marks another step in Chinese labs racing to define the open-weight frontier.