Tencent has officially released and open-sourced Hy4 preview, a next-generation large language model designed for engineering, coding, and general analytical productivity tasks. Built on a Mixture-of-Experts (MoE) architecture, the model scales to 770 billion total parameters while activating 49 billion parameters per token during inference, combined with a context window exceeding 1 million tokens.
Alongside the model release, Tencent outlined details regarding its API availability, cost structure, and an experimental development pipeline where the model participated in optimizing its own training routines and deployment infrastructure.
Sparse Architecture: 770B Total vs. 49B Active Parameters#
The choice of a Mixture-of-Experts architecture allows Hy4 preview to maintain high parameter capacity while controlling computational overhead during inference. Out of its 770 billion total parameters, only 49 billion are activated for any given token routing step.
This parameter distribution aims to balance raw model capacity with operational costs. Supporting a context window of over 1 million tokens, the architecture is targeted at long-context software development, multi-document analysis, and complex reasoning pipelines that exceed the working memory of conventional dense models.
Autonomous Infrastructure and Pipeline Optimization#
A notable technical element reported by Tencent is the model's participation in its own development lifecycle. During pre-training and post-training, Hy4 preview was deployed to propose, execute, and iterate upon training adjustments, data strategies, and evaluation routines, forming an early recursive self-improvement loop.
At the system level, the model was instructed to analyze bottlenecks in its inference stack. Through iterations on operator fusion and communication layer strategies, Tencent reports an end-to-end throughput increase of 31.8% over the baseline system across varying context lengths and concurrency workloads.
Benchmark Performance and Internal Evaluation Context#
To evaluate functional performance, Tencent conducted an internal blind evaluation involving 163 domain experts across 203 engineering tasks. In this evaluation framework, Hy4 preview scored an average of 2.99 out of 4.00, compared to GLM-5.3 (2.92/4.00) and Kimi K3 (2.94/4.00).
While these figures indicate competitive performance in targeted engineering workflows—including code debugging, front-end generation, financial analysis, and scientific problem-solving—the results stem from Tencent's internal benchmark suite. Broader validation on public open-source benchmarks (such as SWE-bench or MMLU-Pro) remains necessary to verify these findings independently.
Pricing, API Access, and Ecosystem Integration#
Hy4 preview has been made available globally across developer tools and cloud infrastructures. Key deployment details include:
- API Pricing: Set at $0.834 USD per million input tokens, $2.501 USD per million output tokens, and $0.042 USD per million tokens for prompt cache hits via Tencent Cloud TokenHub and OpenRouter.
- Platform Access: Integrated into Tencent's productivity and development applications, including WorkBuddy, CodeBuddy, Yuanbao, and ima.
- Trial Period: Offered free of charge for two weeks on the WorkBuddy and CodeBuddy platforms following launch.
Original source: Tencent