Open Frontier Intelligence releases Kimi K3, the first open 3T-class model
The release marks a significant step in open-source large language models, with full weights scheduled for July 2026 and performance claims challenging proprietary incumbents.
Open Frontier Intelligence has released Kimi K3, a 2.8-trillion-parameter large language model described as the world’s first open 3T-class model. The release introduces native multimodal vision capabilities, a 1-million-token context window, and architectural innovations including Kimi Delta Attention and Attention Residuals. The model is currently available via Kimi.com, Kimi Work, Kimi Code, and the Kimi API, with full model weights scheduled for release by 27 July 2026.
The architecture utilises a Stable LatentMoE framework, activating 16 out of 896 experts to achieve approximately 2.5 times the scaling efficiency of its predecessor, Kimi K2. This structural change, combined with refined training and data recipes, allows the model to convert compute into intelligence more effectively. The model also employs quantization-aware training from the supervised fine-tuning stage, using MXFP4 weights and MXFP8 activations for broad hardware compatibility.
In performance evaluations, Kimi K3 demonstrated frontier-level capabilities in coding, knowledge work, and reasoning. In GPU kernel optimisation tests conducted within identical sandboxes, the model substantially outperformed Opus 4.8, GPT 5.6 Sol, and GPT 5.5. It performed competitively with Claude Fable 5, which was evaluated by a third party and may include fallback behaviour. The model also autonomously developed MiniTriton, a compact compiler that delivers performance on par with or better than existing tools like Triton and torch.compile on certain workloads.
Autonomous capabilities extend to hardware design and scientific research. In a 48-hour autonomous run, Kimi K3 designed, optimised, and verified a chip using open-source EDA tools, achieving 100 MHz timing closure and over 8,700 tokens per second decode throughput in simulation. Additionally, the model autonomously reproduced computational astrophysics workflows related to I–Love–Q universal relations, completing in two hours what typically takes researchers one to two weeks.
Deployment recommendations suggest using supernode configurations with 64 or more accelerators to manage expert-parallel training and inference efficiency. To support this, Open Frontier Intelligence has contributed a vLLM implementation to address challenges posed by Kimi Delta Attention regarding conventional prefix caching. The company plans to release further technical details on architecture, training, and evaluations alongside the Kimi K3 technical report.
