Kimi has announced Kimi K3, a 2.8-trillion-parameter multimodal model designed for coding, reasoning and knowledge work. The company made the model available through Kimi.com, Kimi Work, Kimi Code and its API, and said it planned to publish the complete weights by July 27, 2026.
K3 uses a mixture-of-experts design that activates 16 of 896 experts for a given computation. Kimi says the system combines Kimi Delta Attention, Attention Residuals and a Stable LatentMoE framework, producing about 2.5 times the overall scaling efficiency of Kimi K2. It also provides native image capabilities and a context window of one million tokens. Architectural and training details were due to accompany a later technical report.
The company describes K3 as the first open model in the roughly three-trillion-parameter class. It also acknowledges that overall performance remains behind the most capable proprietary systems it tested, while claiming frontier-level results across its own evaluation suite. At launch, the hosted model uses maximum thinking effort by default; lower- and higher-effort choices were planned for future updates.
Kimi highlighted long-running software tasks as a core use case. In company demonstrations, K3 optimized GPU kernels in a 24-hour sandbox and built MiniTriton, a compact compiler with an MLIR-based intermediate layer and PTX generation. Kimi says MiniTriton matched or exceeded Triton and `torch.compile` on supported roofline tests and sustained nanoGPT training with stable convergence. These results are vendor-reported and await broader independent reproduction.
Another demonstration assigned an early K3 build to design and verify a small chip with open-source electronic-design tools. Kimi reported that the simulated design fit within four square millimeters, closed timing at 100 MHz and achieved more than 8,700 decode tokens per second. A separate computational-astrophysics case involved reviewing more than 20 papers, evaluating over 300 equations of state and generating more than 3,000 lines of Python.
Such case studies show the intended direction: agents that can maintain context and use tools across hours rather than answer isolated prompts. The more consequential test will come after weight release, when outside researchers can inspect deployment requirements, reproduce evaluations and assess practical performance. Until then, K3’s scale and availability are established by Kimi’s announcement, while its comparative capabilities remain claims from the developer.
