Chinese AI startup Moonshot AI has achieved a significant breakthrough by unveiling Kimi K3, the largest open-source model in the 3T class to date. With 2.8 trillion parameters, native vision, and a context window of 1 million tokens, this system confidently enters the elite ranks of global AI, trailing only the proprietary giants Claude Fable 5 and GPT 5.6 Sol, according to the developer's own data.
The architectural foundation of K3 is the innovative Kimi Delta Attention (KDA) and the Attention Residuals (AttnRes) mechanism. KDA optimizes the processing of long data sequences, while AttnRes allows the model to selectively extract critically important information without processing all layers equally. The result is a deeper understanding of context and preservation of meaning even when working with the most complex texts. Computational sparsity is ensured by the Stable LatentMoE framework: out of 896 experts, only 16 are active simultaneously, which, combined with new data and training methods, has led to a 2.5-fold increase in efficiency compared to its predecessor, Kimi K2.
Benchmarks: Where K3 Shines and Where It Falls Short
The test results for K3 are mixed. The model demonstrates outstanding performance on tasks requiring long engineering sessions and complex agent behavior. On SWE Marathon, it scored 42 points, surpassing Fable 5 (35) and GPT-5.6 Sol (39). In the BrowseComp test (91.2) and Program Bench (77.8), it also came out ahead. On Terminal Bench 2.1, its result of 88.3 points is only 0.5 points behind Sol (88.8) and significantly exceeds Fable 5 (84.6).
However, on benchmarks assessing deep scientific knowledge, K3 lags behind. On FrontierSWE, it scored 81.2 compared to 86.6 for Fable 5, and on HLE-Full, it scored 43.5 versus 53.3. It is important to note that testing methodologies differed, which may have influenced the final figures.
Coding, Chips, and Creativity: Practical Use Cases
K3 is capable of autonomously conducting lengthy engineering sessions, navigating vast repositories, and managing terminal tools. In one test, the model spent 24 hours optimizing GPU cores, achieving results on par with Fable 5. Moreover, an early version of K3 independently wrote MiniTriton, a compact compiler for GPU cores.
Demonstration cases are impressive: K3 designed a chip for a neural network on its own architecture in 48 hours, using open-source tools. The resulting chip, with an area of 4 mm², delivers over 8700 tokens per second. The model can also create game prototypes, edit videos, and conduct complex scientific analysis, for example, reproducing universal relationships from computational astrophysics in two hours instead of two weeks of researcher work.
Analytical commentary from Cryptalist: Kimi K3 is undoubtedly a step forward for open-source models. However, the gap with proprietary leaders in tasks requiring deep knowledge and flawless usability is still noticeable. Notably, Moonshot AI is betting on autonomy in engineering tasks, which could become their main trump card in the software development niche. But for the mass user, K3 remains a niche, albeit powerful, tool for now.