Kimi K3 releases its weights and open-sources three bricks of its infrastructure

Moonshot AI's Kimi K3 model releases its weights and test sandboxes. What this open infrastructure changes for agent development.

The full weights of Kimi K3, Moonshot AI's 2.8-trillion-parameter open model, are now available as of the announced date of July 27. They are accompanied by a technical report and the open-sourcing of part of the stack that runs the model: FlashKDA, the in-house attention kernels; MoonEP, a communication library for Mixture of Experts; and AgentENV, the sandbox infrastructure designed to run agent environments at scale.

The report reveals several design choices. The vision encoder, MoonViT-V2, was trained from scratch rather than initialized from a pre-existing contrastive model, a choice Moonshot says is more stable during joint training with the language model. The attention mechanism alternates three linear KDA layers for every global layer, without explicit positional encoding, as the position is carried by the recurrent mechanism itself, which allows extension up to one million tokens without modifying the parameters. Furthermore, the context window was extended in stages, from eight thousand to one million tokens, to concentrate the cost of long context on a fraction of the budget.

The infrastructure component is particularly noteworthy. The sandboxes rely on isolated microVMs capable of resuming execution in about fifty milliseconds, pausing without consuming resources while an agent waits for a response, or duplicating themselves to evaluate a path without side effects. To illustrate the scale, Moonshot estimates that more than fifty-one million sandboxes were created during training and evaluation.

Regarding benchmarks, the company points to third-party evaluations. Kimi K3 achieves an intelligence score of 57.1 on Artificial Analysis and 74.7 on the Vals sector index, and takes first place on the WebDev Arena among models ranked by human preference. Moonshot also acknowledges that overall performance remains below that of the most powerful proprietary models, particularly in research-level reasoning.