In brief: Kimi K3 is a mixture-of-experts model with 104 billion active parameters that elevates open-weights models to a new performance tier and has released complete infrastructure tools for large-scale agent training.
Moonshot AI has released the language model Kimi K3 as an open-weights version, significantly closing the performance gap to closed models: the 2.8 trillion parameter model outperforms Claude Opus 4.8 and reduces the distance to leading proprietary systems to four points on the Artificial Analysis Intelligence Index.
Moonshot AI has published Kimi K3 as a complete open-weights package: a mixture-of-experts model with 2.8 trillion total parameters, 104 billion active parameters per token, 896 experts and native image processing. The model features a 1 million token context and was trained with specialized technical optimizations.
Beyond the model weights, Moonshot opened three components of the training and serving infrastructure: FlashKDA (Flash Kimi Delta Attention Kernels), MoonEP (MoE Communication Library) and AgentENV (infrastructure for distributed agents). This goes beyond a standard model release and provides practitioners with a complete blueprint for large-scale agent training and inference.
The technical report documents a scaling efficiency improvement of approximately 2.5× over Kimi K2. Central to this are architectural and training decisions for numerical stability in the extreme range: MXFP4 weights, MXFP8 activations, joint training of the vision encoder from scratch and optimized MoE routing mechanics. Independent evaluations confirm that K3 outperforms Claude Opus 4.8 and thus holds the title of best open-weights model.
For Chief Data Officers: the release places open-weights models for the first time in a performance tier that directly challenges proprietary frontier models. The accompanying infrastructure significantly lowers entry barriers for fine-tuning and agent training on an open basis, while the documented scaling techniques provide guidance for own training processes.
Source: www.latent.space · Published July 28, 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.3.