inference model

On December 6, according to news, yesterday, Google officially launched a new generation of reasoning models, Gemini 3 Deep Think, known as the “IMO gold-class” model, with a significant leap forward from the previous generation, Gemini 2.5 Deep Think, leading in many international benchmarking tests. Gemini 3 Deep Think uses a "parallel reasoning" structure, which simultaneously explores a variety of assumptions, thus demonstrating a breakthrough in complex mathematical, scientific and logical dilemmas. In Hum..

Aug. 7 (Bloomberg) -- Baidu plans to launch a new inference model capable of handling more complex tasks by the end of this month (the end of August 2025) in order to gain a foothold against the likes of DeepSeek and OpenAI, according to a report in today's Wall Street Journal by people with knowledge of the situation. Baidu also plans to release an updated version of its core foundation model, called Ernie 5.0 (note: Ernie is Wenxin), in the coming months, the report added. Baidu released the "strongest" Wenxin model in March ...

July 31, 2011 - Ali Tongyi Qianqian today announced the launch of a new reasoning model Qwen3-30B-A3B-Thinking-2507. Compared to the Qwen3-30-A3B model open-sourced on April 29th, the new model has significantly improved its reasoning ability, generalization ability, and context length: The new model scored 85.0 points in AIME25, a test that focuses on mathematical ability, and 66.0 points in LiveCodeBench v6, a test of code ability. The new model scored 85.0 in AIME25, a math-focused test, and 66.0 in LiveCodeBench v6, outperforming Ge...

Yesterday, Ali Tongyi team announced the official launch of the upgraded version of Qwen3-235B-A22B Thinking model: Qwen3-235B-A22B-Thinking-2507. According to the introduction, the brand new open source Qwen3-235B-A22B-Thinking-2507 realizes a huge leap in inference performance and generalization ability, which can be The new open source Qwen3-235B-A22B-Thinking-2507 achieves a huge leap in inference performance and generalization capability, and can be compared to top closed source models such as Google Gemini-2.5 pro and OpenAI o4-mini, and sets a record for the world's open source model SO...

On June 17, MiniMax announced that it will release important updates for five consecutive days. The first update today is MiniMax-M1, the first open source inference model. According to the official introduction, MiniMax-M1 is the world's first open source inference model with large-scale hybrid architecture, and MiniMax says: M1 is the best open source model in complex productivity-oriented scenarios, surpassing domestic closed-source models and approaching the most advanced models overseas, while offering the industry's best price/performance ratio. The official blog also mentioned that M1 is based on two major technologies. The official blog also mentioned that based on two major technological innovations, Min...

June 11 (Bloomberg) -- French AI lab Mistral announced Tuesday that it is entering the field of inferential AI models. On June 10, Mistral officially launched its first family of inferential models, Magistral, which solves problems in a step-by-step process designed to improve consistency and reliability in disciplines such as math and physics. This series of models is designed to improve consistency and reliability in disciplines such as mathematics and physics through step-by-step problem solving, similar to other inference models such as OpenAI's o3 model and Google's Gemini 2.5 Pro. The Magistral series consists of two versions: Magist...

June 11, 2011 - OpenAI is constantly updating its new generation of generative AI models at a rapid pace. After launching o3 and o4-mini on the ChatGPT platform in April, OpenAI's most powerful model, o3-pro, is now available. The o3 and o4-mini models are designed to integrate with all of ChatGPT's available tools. Models labeled "pro" mean that more time can be devoted to processing questions in exchange for higher quality and more accurate answers. OpenAI's work on the o3-p...

May 14, 2011 - Epoch AI, a nonprofit AI research organization, has released a new report stating that AI companies are struggling to consistently extract significant performance gains from inference models, and that advances in inference models will slow down as soon as a year from now. Based on publicly available data and assumptions, the report highlights computational resource constraints and increased research overheads.The AI industry has long relied on these models to improve benchmark performance, but that reliance is being challenged. Josh You, an analyst at the agency, points to the rise of inference models, which stem from their ability to excel at specific tasks. For example, OpenA...

April 18, 2011 - More and more users are using ChatGPT to decipher the exact location where a photo was taken, a new and worrying phenomenon that is rapidly spreading across the web, according to a report by the foreign media outlet TechCrunch today. This week, OpenAI launched two new models -- o3 and o4-mini -- with image inference capabilities that can analyze details in uploaded photos, and even crop, rotate, and zoom in on blurry or distorted images for deeper recognition. With this analysis capability, combined with the model's web search function...

April 17th, OpenAI officially released the o3 and o4-mini models in a live broadcast early this morning.1AI summarized the content as follows: OpenAI's official introduction said that this is the result of its latest training in the o-series of models, which can be used to think for longer periods of time before answering, and also claimed that it is "the most intelligent model OpenAI has released so far". It also claims to be "the smartest model OpenAI has released to date" and represents a major leap forward in ChatGPT's capabilities that will benefit everyone from curious users to advanced researchers. For the first time, the new inference model can intelligently use and combine...

1AI learned from Tencent mixed yuan WeChat public number, March 21 Tencent mixed yuan officially launched the official version of the self-developed deep thinking model mixed yuan T1. According to the official introduction, "T1" spits out words fast, can return in seconds, but also good at ultra-long text processing, is Tencent's self-research strong reasoning model. Through large-scale reinforcement learning, and combined with special optimization of mathematics, logical reasoning, science and code and other science challenges, the official version of Mixed Elements T1 further improves its reasoning ability. On common benchmarks that reflect the basic capabilities of inference models, such as the large language model evaluation enhancement dataset MMLU-PRO,...

On March 19, Tencent announced via its official public website that its new reasoning model, T1, will be officially released this Friday, March 21 at 23:00 BST. At the same time, Tencent announced that for the first time, the hybrid model has made it onto the Chatbot Arena list, ranking among the top 15 in the world, where users anonymously interact with multiple models to vote on which is better, generating a leaderboard based on their scores. This kind of evaluation is also seen as a direct PK arena for big models, which is simple and straightforward. On September 5 last year, Tencent...

March 6, 2011 - Research has shown that reinforcement learning can significantly improve the inference capabilities of models such as DeepSeek-R1, which achieves state-of-the-art performance by integrating cold-start data and multi-stage training, enabling deep thinking and complex inference. AliCloud Tongyi Qianqi officials today announced the availability of their latest inference model, QwQ-32B, a model with 32 billion parameters that rivals the performance of DeepSeek-R1 with 671 billion parameters, 37 billion of which are activated. This achievement highlights the importance of bringing reinforcement learning to ...

March 4, 2011 - Amazon is developing an artificial intelligence model with advanced "reasoning" capabilities, Business Insider reported this evening. According to a source directly involved in the project, the new product is scheduled to be released as early as June under the brand name "Nova," a series of generative AI models that Amazon launched late last year. The person said Amazon wants the new models to use a "hybrid reasoning" approach, which can give quick answers, but also more complex thinking in the same system. Reasoning models have become an AI field in recent years...

February 19th news, tencent meta official tonight announced that its deep thinking model "mixed yuan T1" can now be used by all users unlimited (mixed yuan + DeepSeek two models free unlimited). According to the introduction, Tencent mixed yuan T1 and DeepSeek-R1 are both reasoning models, can understand the multiple dimensions of the problem and the underlying logical relationships, especially suitable for completing complex tasks. In addition, Tencent Yuanbao can not only use DeepSeek-R1 full-blooded version and Mixed Yuan T1 for deep thinking, but also can use DeepSeek-...

Feb. 6 (Bloomberg) -- Artificial intelligence researchers at Stanford University and the University of Washington have successfully trained an AI model with "reasoning" capabilities for less than $50 (note: currently about 364 yuan) in cloud computing, according to a study released on Friday. The model, called s1, performed similarly to top reasoning models such as OpenAI's o1 and DeepSeek's r1 in tests of mathematical and programming ability. The s1 model and the data and code used to train it are now available on GitHub ...

January 22 news, NetEaseYouDao today announced the launch of the first domestic output step-by-step explanation of the reasoning model "ZiYao-o1", and officially open source. According to the official introduction, ZiYao-o1 is a 14B lightweight single model, support in the deployment of consumer-grade graphics cards, the use of chain of thought technology, able to provide a detailed problem solving process, to strong logic and reasoning ability, to achieve higher problem solving accuracy, and to provide the Chinese logical reasoning. NetEase said that there are not many open source models available for deployment, and the parameter scale is large, which can't be run on consumer-grade graphics cards with low graphics memory, even with low-bit quantization technology, making...

December 18, 2012 - Entering the ninth day of its "12 Days of OpenAI" campaign, OpenAI today announced that its "inference" AI model, o1, has been officially opened to some developers via APIs, and has been updated with a number of developer tools, including GPT-4o, real-time APIs, and fine-tuning APIs. OpenAI today announced that its "inference" AI model o1 is officially available to some developers via APIs, with simultaneous updates to developer tools including GPT-4o, real-time APIs and fine-tuning APIs. It is reported that the first batch of developers who can use the o1 API are OpenAI's "Level 5" users. To reach this level, developers need to spend at least $1,000 on the OpenAI platform (...

According to Chengdu Municipal Bureau of Science and Technology, on November 22, Chengdu Humanoid Robotics Innovation Center made a breakthrough in technological innovation, and recently launched the first diffusion-based high-speed reasoning model for humanoid robots, R-DDIRM (Denoising Diffusion Implicit Robot Model). Following the launch of China's first diffusion-based humanoid robot task generative model R-DDPRM (Denoising Diffusion Probability Robot Model) in May this year, Chengdu Humanoid Robotics Innovation Center has made a breakthrough in technological innovation.

November 21st, DeepSeek announced that the preview version of its newly developed inference model DeepSeek-R1-Lite is officially online. Officially, the DeepSeek R1 series of models are trained using reinforcement learning, and the reasoning process includes a great deal of reflection and verification, and the chain of thought can be tens of thousands of words long. The series of models have achieved reasoning results comparable to OpenAI o1-preview in math, code, and a variety of complex logical reasoning tasks, and have shown users the complete thinking process that o1 has not disclosed. DeepSe...