[{"_1":2,"_100":-5,"_101":-5},"loaderData",{"_3":4,"_6":7},"root",{"_5":-5},"user","routes/tag.$slug",{"_8":9,"_89":90,"_97":92,"_98":62,"_99":62},"posts",[10,70],{"_11":12,"_13":14,"_15":16,"_17":18,"_19":20,"_21":22,"_23":24,"_25":26,"_27":28,"_29":30,"_31":32,"_33":-5,"_34":35,"_36":37,"_47":48,"_63":64,"_65":66,"_67":68,"_69":32},"documentId","ezo5pkvfi3ubfn12wybouloc","title","Tencent Hunyuan's New Generation Turbo S Fast Thinking Model Released","slug","tencent-hunyuans-new-generation-turbo-s-fast-thinking-model-released","excerpt","On February 27th, Tencent's Hunyuan new generation fast thinking model Turbo S was officially released.","content","\n

On February 27th, Tencent's Hunyuan new generation fast thinking model Turbo S was officially released.

\n\n\n\n

Tencent officials stated that unlike the slow thinking models such as Deepseek R1 and Hunyuan T1 that require \"thinking before answering,\" Hunyuan Turbo S can achieve \"instant response,\" output answers more quickly, double the speaking speed, and reduce initial latency by 44%.

\n\n\n\n

In several widely used benchmarks in the industry, Hunyuan Turbo S has shown performance comparable to a series of business models such as DeepSeek V3, GPT 4o, Claude in various fields including knowledge, mathematics, reasoning.

\n\n\n\n

In terms of architecture, the Hunyuan Turbo S adopts the Hybrid-Mamba-Transformer fusion mode, reducing the computational complexity of traditional Transformer structures and decreasing KV-Cache cache usage, achieving a decrease in training and inference costs.

\n\n\n\n

Tencent also stated that the new fusion mode breaks through the difficulty faced by large models with pure Transformer structures in high training and inference costs for long texts. On one hand, it leverages Mamba's efficient processing of long sequences; on the other hand, it retains the advantage of Transformers in capturing complex contexts, ultimately building a hybrid architecture with dual advantages in memory and computational efficiency.

\n\n\n\n

This is also the first successful application of Mamba architecture without loss on super-large MoE models in industry. Through technical innovation in model architecture, Hunyuan Turbo S significantly reduces deployment costs.

\n\n\n\n

As a flagship model, Hunyuan Turbo S will become the core foundation for Tencent's derived models from the Hunyuan series in providing basic capabilities for inference, long texts, code generation, and other derived models.

\n\n\n\n

Based on Turbo S and introducing technologies such as long thought chains, retrieval enhancement, and reinforcement learning，Hunyuan has also launched an inference model T1 with deep thinking capabilities.

\n\n\n\n

From the Tencent announcement, it is learned that currently, developers and enterprise users can now call Tencent Hunyuan Turbo S through API on Tencent Cloud for a free trial within one week.

\n\n\n\n

In terms of pricing, the input price for Turbo S is 0.8 yuan per million tokens, and the output price is 2 yuan per million tokens.

\n\n\n\n

Tencent Yuanbao will gradually launch Hunyuan Turbo S in grayscale. Users can experience using it by selecting the \"Hunyuan\" model in Yuanbao and turning off deep thinking.

\n\n\n\n

\n","publishedDate","2025-02-28T14:24:17.000Z","featured",false,"viewCount",0,"readingTime",3,"status","draft","language","en","featuredImage","featuredImageUrl","https://assets.pandaily.com/uploads/2023/04/tc.jpeg","categories",[38],{"_11":39,"_40":41,"_15":42,"_43":44,"_45":46},"ou7tkn9cq6eba9jr68hqiwch","name","AI","ai","color","#000000","description","","tags",[49,54,56,60],{"_11":50,"_40":51,"_15":52,"_43":53,"_45":46},"njwm066t5dgnkyahw9o23jvu","Tencent","tencent","#6366f1",{"_11":55,"_40":41,"_15":42,"_43":53,"_45":46},"ji5tef72gjtbd7bx0er6q9jv",{"_11":57,"_40":58,"_15":59,"_43":53,"_45":46},"j45fwu61bg7hx57fo10s37oq","China","china",{"_11":61,"_40":62,"_15":62,"_43":53,"_45":46},"vkmskg5gqw144f9z2dkqbmbe","hunyuan","createdAt","2025-06-24T22:14:18.934Z","updatedAt","2025-06-24T23:44:54.240Z","publishedAt","2025-06-24T23:44:54.275Z","locale",{"_11":71,"_13":72,"_15":73,"_17":74,"_19":75,"_21":76,"_23":24,"_25":26,"_27":77,"_29":30,"_31":32,"_33":-5,"_34":78,"_36":79,"_47":81,"_63":86,"_65":87,"_67":88,"_69":32},"qzwlxoa83fhwkv2hgub4z4st","Tencent Releases Open-Source MoE Large Language Model Hunyuan-large","tencent-releases-open-source-moe-large-language-model-hunyuan-large","Tencent released the open-source MoE large language model Hunyuan-large with a total of 398 billion parameters, making it the largest in the industry.","\n

On November 5th, Tencent released the open-source MoE large language model Hunyuan-large with a total of 398 billion parameters, making it the largest in the industry, with 52 billion activation parameters. Public evaluation results show that Tencent's Hunyuan Large model leads comprehensively in various disciplines such as CMMLU, MMLU, CEval, MATH comprehensive evaluation sets and tasks including NLP in Chinese and English languages, code processing, mathematics across nine dimensions surpassing top-tier open-source large models like Llama3.1 and Mixtral. It is understood that this model can achieve high-quality synthetic data through technical innovation by using synthetic data to enhance training effectively addressing the shortcomings of natural data. In terms of context processing capability, the pre-trained model supports text sequences up to 256K significantly enhancing its ability to handle long-context tasks.

\n\n\n\n

At the same time, Tencent Hunyuan announced that to address the lack of real-world long-text evaluation sets in the industry they will soon release an open-source PenguinScrolls evaluation set to support application research within the industry. The self-developed PenguinScrolls is based on various public financial documents, legal texts academic papers among other types of natural long texts ranging from lengths between 1K-128K covering various deep reading comprehension and long-text reasoning tasks.

\n\n\n\n

\n","2024-11-06T21:13:59.000Z",2,"https://assets.pandaily.com/uploads/2024/05/TENCENT.jpg",[80],{"_11":39,"_40":41,"_15":42,"_43":44,"_45":46},[82,85],{"_11":83,"_40":51,"_15":84,"_43":53,"_45":46},"wf5fs0kpm2w6xoqqr718rix1","tencent-ms",{"_11":61,"_40":62,"_15":62,"_43":53,"_45":46},"2025-06-24T22:16:55.644Z","2025-06-24T23:44:53.200Z","2025-06-24T23:44:53.233Z","pagination",{"_91":92,"_93":94,"_95":92,"_96":77},"page",1,"pageSize",12,"pageCount","total","currentPage","tag","tagSlug","actionData","errors"]

Tag: #hunyuan

Tencent Hunyuan's New Generation Turbo S Fast Thinking Model Released

Tencent Releases Open-Source MoE Large Language Model Hunyuan-large