Tencent releases and open-sources Hy4 preview: 770B parameters, 1M-token context

Tencent released and open-sourced Tencent Hy4 preview on 28 August — a next-generation large language model with 770 billion total parameters, 49 billion active parameters and a context window exceeding one million tokens. The company says it is built for real-world productivity tasks across coding, office work and scientific research.

Availability and pricing

Hy4 preview is available as an open-source model and can be accessed globally through WorkBuddy, CodeBuddy, Yuanbao, ima and other Tencent products, or connected by API via Tencent Cloud TokenHub and OpenRouter. On launch it is free on WorkBuddy and CodeBuddy for two weeks, and free access to Hy3 on both platforms has been extended until 30 September. API pricing is set at USD 0.834 per million input tokens, USD 2.501 per million output tokens and USD 0.042 per million tokens for cache hits.

Internal evaluation and capabilities

In an internal blind evaluation involving 163 experts and 203 engineering tasks, Hy4 preview scored an average of 2.99 out of 4.00, slightly ahead of GLM-5.3 (2.92) and Kimi K3 (2.94). The model was trained on high-quality data co-created with Tencent experts in software engineering, gaming, finance and security. In software engineering it handles long-context development tasks better, improving understanding, planning, debugging and validation, along with front-end visual quality; in office and analytical work it strengthens financial analysis and cross-document collaboration, supporting documents, spreadsheets and presentations; in game development it can generate a playable prototype from a single natural-language request; and in research it shows gains in AI research and development, molecular dynamics simulation, condensed-matter physics and fundamental mathematics.

Self-improvement and inference gains

Notably, Hy4 preview took part in its own development for the first time, contributing to the automated optimisation of training methods, data strategies, evaluation frameworks and low-level operators. It proposed approaches, ran experiments and iterated on the results, feeding code, logs and feedback into later rounds — what Tencent describes as an early-stage recursive self-improvement loop. The model also analysed bottlenecks in its own inference system and carried out several rounds of optimisation on operator fusion and communication, increasing end-to-end throughput by 31.8% over the baseline, with consistent gains across context lengths and concurrency levels. Tencent says the next batch of models in the Hy4 series is expected soon.