Tencent's new-generation fast-thinking hybrid model Turbo S released, supports "second reply"

What is the Hybrid Turbo S

The Hybrid Turbo S isTencent hybridOn February 27, 2025, the self-developed new generation of fast thinking model was officially released. The model is designed to solve the shortcomings of the slow thinking model in response speed, and realize the "second reply" capability through technological innovation, doubling the speed of spitting and reducing the delay of the first word by 44%.

Adopting the Hybrid-Mamba-Transformer fusion model, Hybrid Turbo S applies the Mamba architecture losslessly to very large MoE models for the first time, which reduces computational complexity, reduces KV-Cache cache occupancy, and significantly lowers training and inference costs. This innovation enables Hybrid Turbo S to have a lower deployment threshold while maintaining high performance.

On a number of publicly available benchmarks common to the industry, the Hybrid Turbo S scoresKnowledge, math, reasoningIt has demonstrated the performance of a series of industry models such as DeepSeek V3, GPT 4o, Claude, etc. in many fields.

Hybrid Turbo S Core Features

1. spontaneous recovery capability::

   • rapid responseThe hybrid Turbo S is able to realize the "second reply", doubling the speed of spitting, reducing the delay of the first word by 44%, and almost realizing the "immediate question and answer", which greatly improves the smoothness of the interaction.
   • wide application: Whether it's a daily conversation, code generation or intelligent customer service scenarios, Hybrid Turbo S delivers a silkier interaction experience and reduces user wait time.

2. superior performance::

   • Strong intellectual, mathematical, and creative skills: Demonstrate effective performance against industry-leading models such as DeepSeek V3, GPT-4o, Claude 3.5, and others in multiple domains of knowledge, math, and reasoning.
   • Integration of short and long term thinking chains: Through the fusion of long and short thought chains, the overall effectiveness of the model is improved by significantly improving science reasoning while maintaining the experience of thinking fast on liberal arts-type problems.

3. Cost optimization::

   • Architecture Innovation: The Hybrid-Mamba-Transformer fusion model is adopted, which effectively reduces the computational complexity and KV-Cache cache occupancy of the traditional Transformer structure, and realizes the reduction of training and inference costs.
   • Lower deployment costs: This innovation has led to a significant reduction in the cost of deploying the Hybrid Turbo S, helping to drive down the barriers to large model adoption.

Hybrid Turbo S Application Scenarios

Hybrid Turbo S is suitable for scenarios that require fast response and efficient processing power, such as intelligent customer service, dialog systems, and code generation. Its efficient and low-cost features make it able to meet the needs of enterprises and developers for highly efficientAI macromodelThe needs of the

Hybrid Turbo S Market Positioning

Hybrid Turbo S asTencent hybridThe flagship model of the new generation of the series is dedicated to providing users with smarter and more efficient AI services. Its API call pricing is $0.8/million tokens for input and $2/million tokens for output, which is several times lower than the price of the previous generation of models and has a higher cost-effectiveness.

How to use Hybrid Turbo S

• Tencent Cloud API Calls: Developers and enterprise users can call hybrid Turbo S via API on Tencent Cloud's official website for a free trial within a week from now. Application address:https://cloud.tencent.com/apply/p/i2zophus2x8
• Tencent Yuanbao ExperienceTencent Yuanbao will soon gradually gray-scale on-line hybrid Turbo S, users in the Yuanbao to select the "Hunyuan" model and close the depth of thinking can experience the use of.