Hunyuan-TurboS

Hunyuan-TurboS was unveiled by Tencent on February 27, 2025, as a 'fast-thinking' flagship positioned against DeepSeek-R1. It is the industry's first ultra-large Hybrid-Transformer-Mamba mixture-of-experts model, combining Mamba's efficient long-sequence processing with Transformer contextual understanding while drastically reducing KV-cache requirements. Tencent claimed 44% faster response times and double the word-generation speed of DeepSeek-R1, with answers in under a second.

On benchmarks including MMLU, C-Eval, MATH, and AIME 2024, Tencent reported performance on par with DeepSeek-V3, GPT-4o, and Claude, at API pricing of roughly $0.11 per million input tokens. It serves as the base for Tencent's Hunyuan-T1 reasoning model and is available via Tencent Cloud APIs only.