December 16th.Dark Side of the Moon Kimi Today's Releasesvisual thinking model k1. The model is built on reinforcement learning techniques and natively supports theEnd-to-End Image Understanding and Chain of Thought Technology, and expanding capabilities to more basic sciences beyond math.
Dark Side of the Moon officials say that the first-generation k1 model outperformed OpenAI o1, GPT-4o, and Claude 3.5 Sonnet in benchmark proficiency tests in basic science disciplines such as math, physics, and chemistry.
1AI has learned from Dark Side of the Moon officials that theKimi's new model will be online as soon as it's released.The k1 Visual Thinking Model is now available in the latest version of Kimi Intelligent Assistant on Android and iPhone mobile apps and on the web at kimi.com..Find "Kimi Visual Thinking Edition" in the latest version of the mobile app or on the Kimi+ page on the web, and you can take a photo or send a picture to experience it.
"Kimi Visual Thinking Edition" will present the complete chain of deductive thinking CoT.Let the user not only see the results of the answer, but also see the whole process of the model to think about the answer.
From the perspective of model training, the training of the k1 visual thinking model is divided into two phases, theThe base model is first obtained by pre-training and then trained on the base model after reinforcement learning.The base model of k1 focuses on optimizing the character recognition capability, obtaining a (state-of-the-art) result of 903 on OCRBench, and scores of 69.1, 66.7, and 96.9 on the MathVista-testmini, MMMU-val, and DocVQA benchmarking sets, respectively.
The Dark Side of the Moon says that k1's reinforcement learning post-training has been further optimized in terms of data quality and learning efficiency, and new breakthroughs have been made in the scaling (scaling) of reinforcement learning.
In addition, a scientific benchmarking program for modeling capabilities is one of the important challenges facing the large modeling industry. Due to the lack of graphical test sets for basic science subjects in the market, Kimi model R&D team has independently constructed a standardized test set Science Vista, which covers mathematical, scientific and chemical graphical topics of different levels of difficulty, and matches the actual users' needs in terms of distribution.The test set will be open to the entire industry and users can apply to use it under license.
In internal testing, Dark Side of the Moon also found some limitations of the k1 visual thinking model, such as generalization of out-of-distribution, success rate on more complex problems, accuracy in more noisy scenarios, and multi-round quizzing effects, which have a lot of room for improvement.In some scenarios and generalization capabilities, the k1 model still falls short of OpenAI's o1 family of models.