On March 21, Tencent announced the official launch of its self-developed deep thinking model, Hunyuan T1, which is now available on the Tencent Cloud official website. The input price is 1 yuan per million tokens, and the output price is 4 yuan per million tokens.
Tencent Hunyuan summarized that the characteristics of the official version of Hunyuan T1 are &#34;fast articulation, instant response&#34; and &#34;good at handling long texts,&#34; claiming it is Tencent's self-developed strong reasoning model.
In terms of speed, the official version of Hunyuan T1 has decoding performance that is twice that of the industry under comparable parameter quantities, with the first word output in seconds and an articulation speed of 60 to 80 tokens per second, demonstrating excellent long text processing capabilities. Multiple indicators reach the level of leading reasoning models in the industry, and in terms of effectiveness, the official version of Hunyuan T1 exhibits strong logic, concise writing style, accurate following of complex instructions, and low hallucination in summaries.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/0d3f8791-7148-451a-9a42-75443ee27053.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/0d3f8791-7148-451a-9a42-75443ee27053.jpeg"/>
Through large-scale reinforcement learning, combined with specialized optimizations for scientific and mathematical problems such as mathematics, logical reasoning, science, and coding, the official version of Hunyuan T1 further enhances its reasoning capabilities.
In common evaluation metrics that reflect the basic capabilities of reasoning models, such as the enhanced dataset MMLU-PRO for large language model (LLM) evaluation, Hunyuan T1 scored 87.2, second only to OpenAI o1's score of 89.3, and higher than OpenAI GPT 4.5's score of 86.1 and DeepSeek R1's score of 84.
In public benchmark tests for Chinese and English knowledge and competition-level mathematics and logical reasoning, such as CEval, AIME, and Zebra Logic, Hunyuan T1's performance also reached the level of leading reasoning models in the industry, with a logical reasoning score of 93.1, surpassing the aforementioned models o1, GPT 4.5, and R1.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/aacf034e-a4a5-462b-a79d-d567c613fde4.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/aacf034e-a4a5-462b-a79d-d567c613fde4.jpeg"/>
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/c23fe079-ebc8-4bf6-9a83-e41fbef9d2e9.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/c23fe079-ebc8-4bf6-9a83-e41fbef9d2e9.jpeg"/>
The powerful reasoning capability of the official version of Hunyuan T1 stems from Tencent's use of the architecture of Hunyuan Turbo S, adopting a Hybrid-Mamba-Transformer fusion model. This is the first time in the industry that the hybrid Mamba architecture has been applied losslessly to ultra-large reasoning models.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/2078d4f5-18fd-4671-93fe-6fd40846acdb.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/2078d4f5-18fd-4671-93fe-6fd40846acdb.jpeg"/>
The aforementioned architecture effectively reduces the computational complexity of traditional Transformer structures and decreases the memory usage of KV-Cache, thereby significantly lowering training and reasoning costs
The Hunyuan T1 also demonstrates unique advantages in the field of long text reasoning.
Based on its outstanding ability to capture long texts, the Hunyuan T1 can effectively address common issues in long text reasoning, such as context loss and long-distance information dependency. At the same time, the hybrid Mamba architecture has been specifically optimized for long sequence processing, significantly reducing resource consumption while ensuring the ability to capture long text information through efficient computation methods, achieving a 2-fold increase in decoding speed with a similar number of activation parameters.
Before launching the new reasoning model, on Monday, Tencent's Hunyuan model made its debut on the overseas authoritative large model competition platform Chatbot Arena, ranking among the global Top 15. Also on the list is Alibaba's latest open-source model QwQ-32B, released on the 6th of this month.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/453a5965-9564-47d5-bd83-fc7fad350917.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/453a5965-9564-47d5-bd83-fc7fad350917.jpeg"/>
Unlike other evaluations, the evaluators of Chatbot Arena are end-users—users interact anonymously with multiple models on the platform and then vote to determine which model is better, generating a leaderboard based on the scores. This evaluation is also seen as an arena for direct competition among large models, simple and straightforward.
The Tencent Hunyuan model also performed outstandingly in the Chinese foundational capability test. In the &#34;Chinese Large Model Evaluation Benchmark SuperCLUE March Report&#34; released on Tuesday, Tencent Hunyuan ranked second among foundational models domestically, with comprehensive strength firmly in the top tier of domestic large models.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/babf81aa-6731-49dc-9290-52ea8f979d73.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/babf81aa-6731-49dc-9290-52ea8f979d73.png"/>

TCTZF

TCEHY

00700

The official version of Hunyuan T1 adopts the Hunyuan Turbo S architecture, marking the first lossless application of the hybrid Mamba architecture in the industry for ultra-large inference models. It boasts decoding performance that is twice that of the industry under comparable parameter counts, with first-word response times and a speaking speed of 60 to 80 tokens per second, excelling in ultra-long text processing. In public benchmark tests that reflect the foundational capabilities of inference models, Hunyuan T1 achieved industry-leading levels, scoring 93.1 in logical reasoning tests, surpassing OpenAI's o1, GPT 4.5, and DeepSeek's R1

- Tencent announced the official launch of its self-developed deep thinking model, Mixyuan T1, on March 21, with pricing set at 1 yuan per million tokens for input and 4 yuan for output.  
- The Mixyuan T1 model boasts superior decoding performance, achieving speeds of 60 to 80 tokens per second, and excels in long text processing, scoring 87.2 in the MMLU-PRO evaluation, just behind OpenAI's model.  
- Utilizing the Hybrid-Mamba-Transformer architecture, Mixyuan T1 reduces computational complexity and memory usage, enhancing its reasoning capabilities and establishing its position among the top models in the industry.

Wallstreetcn

Tencent Holdings Limited

Tencent

TENCENT

OpenAI

BABA-W

The official version of Tencent's Hunyuan Deep Thinking Model T1 is here, with fast articulation, instant responses, and a decoding speed increase of 2 times