Moonshot AI, the Beijing-based lab behind the Kimi chatbot, published the open weights and technical report for Kimi K3 this week, closing out a release window the company set on July 16. The headline in that earlier announcement was the model itself: a 2.8 trillion-parameter Mixture-of-Experts (MoE) system, an architecture that activates only a slice of its total parameters for each token instead of running the whole network every time. What shipped this week is different. Alongside the weights and the paper, Moonshot released a set of high-performance attention kernels, an MoE communication library, and infrastructure for running agent environments at scale.
Publishing weights and calling a model open are not the same as making it usable. A model with 2.8 trillion total parameters, even one that activates a small fraction of them per token, still has to be held in memory across a cluster of GPUs, with those active experts routed to and from the right hardware in milliseconds. Without software built to do that efficiently, a downloaded checkpoint is closer to a research artifact than a product a team can actually serve.
The attention kernels are built around Kimi Delta Attention, the mechanism Moonshot introduced with K3 to speed up decoding across long contexts. The company says the technique can cut decoding time by up to 6.3 times at the million-token context length K3 supports. Releasing the kernel code, rather than describing the method in a paper alone, lets other teams run that speedup instead of reverse-engineering it themselves.
The MoE communication library targets a separate bottleneck: routing tokens to the correct experts across many GPUs without network traffic becoming the slowest part of the system. Kimi K3 activates 16 of 896 experts per token through what Moonshot calls a Stable LatentMoE framework, a routing scheme wide enough that coordinating it efficiently across a data center is its own engineering problem. A public communication library turns that problem into a solved one for anyone deploying the model, rather than something each team rebuilds from scratch.
The third piece, infrastructure for running agent environments at scale, points at what Moonshot built K3 to do. The model targets long-horizon agentic coding and what the company calls self-evolving workflows, tasks where an agent acts, observes the result, and acts again across many steps rather than answering a single prompt. Opening the environment infrastructure gives outside developers the scaffolding to test and train agents against K3 without building that harness themselves.
Moonshot credits the model’s efficiency to Kimi Delta Attention combined with Attention Residuals, an architectural change the lab published separately in March, plus a larger and sparser pool of experts. Together, the company says, these changes give K3 roughly 2.5 times the scaling efficiency of its predecessor, Kimi K2, meaning the model converts a given amount of compute into more capability. That 2.5x figure is Moonshot’s own claim, measured against its own prior model, not an independently verified benchmark.
For teams evaluating open-weight alternatives to closed frontier models, the question this week is not whether Kimi K3 is capable. Moonshot made that case on July 16. The open question is whether the released kernels and communication library hold up under real multi-node deployment. Teams with the GPU budget to run a 2.8 trillion-parameter model should benchmark Moonshot’s released infrastructure against their existing MoE serving stack before committing engineering time to a migration.
Moonshot AI announced the Kimi K3 release, including its open weights and supporting infrastructure, in a thread posted to X on July 16, 2026.
