Recently, Beijing Moonlit Dark Side Technology Co., Ltd. announced a significant technological upgrade for its intelligent assistant Kimi, introducing the new k1.5 multimodal thinking model. This model achieves industry-leading levels in multimodal reasoning and general reasoning capabilities, marking another breakthrough for Kimi in the field of artificial intelligence.
The k1.5 multimodal thinking model is the third major upgrade of Kimi's k series reinforcement learning models within just three months. Following the release of the k0-math mathematical model in November last year and the k1 visual thinking model in December, the k1.5 model has performed excellently in benchmark tests. In the short-CoT mode, k1.5's mathematical, coding, visual multimodal, and general capabilities significantly surpass the levels of globally recognized short reasoning SOTA models GPT-4o and Claude3.5Sonnet by as much as 550%. In the long-CoT mode, k1.5's mathematical, coding, and multimodal reasoning capabilities also reached the level of the long reasoning SOTA model OpenAI o1 official version, making it the first company outside of OpenAI to achieve multimodal reasoning performance equivalent to the o1 official version.
This upgrade is the result of the relentless efforts and innovations of the Kimi technical team. The team has publicly released a detailed model training technical report titled "Kimi k1.5: Scaling Reinforcement Learning with Large Language Models," documenting the exploration of model training under the new technological paradigm.
The report highlights key innovations of the k1.5 model, including long context extension, which improves training efficiency through partial unfolding techniques, and notes that increasing context length can continuously enhance model performance. Additionally, improved strategy optimization methods and a streamlined framework design support the model's strong performance. Notably, the k1.5 model was jointly trained on text and visual data, enabling joint reasoning across both modalities, particularly excelling in mathematical capabilities, although challenges remain in handling geometry problems that rely on graphic understanding.
To further enhance short-chain reasoning capabilities, the team also proposed an effective long2short method, utilizing Long-CoT technology to improve the Short-CoT model, achieving significant results in tests such as AIME, MATH500, and LiveCodeBench, far surpassing existing short-chain reasoning models like GPT-4 and Claude Sonnet3.5.
The preview version of the k1.5 multimodal thinking model will gradually roll out on the Kimi.com website and the latest version of the Kimi intelligent assistant app. Users can experience this newly upgraded model by finding the model switch button during use. The k1.5 model excels in deep reasoning, helping users solve complex coding issues, mathematical problems, and work-related challenges.
Moonlit Dark Side Technology Co., Ltd. stated that in 2025, it will continue to accelerate upgrades to the k series reinforcement learning models along the established roadmap, bringing more modalities, capabilities in more fields, and stronger general capabilities to unlock more possibilities for users.
GitHub report link: https://github.com/MoonshotAI/kimi-k1.5
