Moonshot AI Releases Kimi K3 Open Weights

Moonshot AI describes Kimi K3 as using Kimi Delta Attention (KDA), Attention Residuals (AttnRes) and a Stable LatentMoE design. The company says the architecture combines 69 KDA attention layers with 24 Gated MLA layers, and uses MXFP4 weights and MXFP8 activations through quantization-aware training.

The model card claims that KDA, AttnRes, a sparser MoE configuration and revised training recipes provide an approximately 2.5x improvement in scaling efficiency over Kimi K2. That is a company-reported efficiency claim, not an independently reproduced measurement.

The Decoder reports that the release also includes a technical report and infrastructure components for attention kernels, MoE communication and large-scale agent operation. Neither that report nor the other retrieved coverage independently reproduces Moonshot AI's full benchmark table, which mixes company-run evaluations with cited third-party leaderboard results and different agent harnesses.

Moonshot AI positions Kimi K3 for long-horizon coding, tool use, knowledge work and multimodal tasks. Those are intended-use claims from the publisher. The immediately verifiable facts are the published weights, sparse-MoE topology, long-context configuration and low-precision format.

Open-weight releases at this scale expand opportunities for inference-engine optimization, expert-parallel serving research, quantization experiments and evaluation on private workloads. In practice, total storage, accelerator memory, communication overhead and hardware support for MXFP4 determine whether the model is feasible to run rather than merely available to download.

Kimi K3 is notable because it makes an unusually large model available under disclosed license terms. Teams evaluating it still need to read that license, reproduce capability claims on their own tasks and match serving requirements to their accelerator, storage and interconnect capacity.