Hy 3d 3.1

The Hy3 official version is honed for real-world business scenarios, using a Mixture-of-Experts (MoE) architecture with 295B total parameters and 21B activated parameters. It natively supports a 256K context window and offers multiple thinking modes: no_think (ultra-fast response), think_low (quick thinking), and think_high (deep reasoning), balancing ultra-fast responses, complex reasoning, and invocation cost. Compared with the Preview version, Hy3—based on real business feedback from Tencent Yuanbao, WorkBuddy, ima, Marvis, and others—focuses on improving the Coding Agent, long-form understanding, multi-turn context continuity, search QA, and complex task execution, performing more stably in reducing hallucinations, improving task completion, and engineering usability. It is better suited to practical scenarios such as frontend tasks, cross-file code development, long-document analysis, office automation, and multi-step Agent workflows.

Hunyuan Hy3 preview is designed for agent workloads, adopting a MoE architecture with 295B capacity and 21B activated parameters. It provides three modes within the same model—no_think (ultra-fast response), think_low (fast thinking), and think_high (deep reasoning)—to accommodate different latency and depth requirements from high-frequency interactions to complex engineering tasks. On code benchmarks such as SWE-bench Verified it approaches the current state of the art, and its 256K context supports cross-file code refactoring and long-document analysis. It is suitable for developers who require reliable task completion while being sensitive to inference costs.

Hunyuan-A13B-Instruct has 8 billion parameters and can match larger models by activating only 1.3 billion parameters, supporting "fast thinking/slow thinking" hybrid inference. It offers stable long text understanding. Verified by BFCL-v3 and τ-Bench, its Agent capabilities are leading in the field. Combined with GQA and multiple quantization formats, it enables efficient inference.

Hunyuan-MT-7B is a lightweight translation model with 7 billion parameters, designed to translate source text into target languages. The model supports translation among 33 languages as well as 5 Chinese minority languages. In the WMT25 International Machine Translation Competition, Hunyuan-MT-7B achieved first place in 30 out of 31 language categories it participated in, demonstrating its exceptional translation capabilities. For translation scenarios, Tencent Hunyuan proposed a complete training paradigm from pre-training to supervised fine-tuning, followed by translation reinforcement and ensemble reinforcement, enabling it to achieve industry-leading performance among models of similar scale. The model is computationally efficient, easy to deploy, and suitable for various application scenarios.