On July 16, 2026, Moonshot AI announced Kimi K3, a large language model. It is a Mixture-of-Experts model with 2.8 trillion total parameters, and a defining feature is its rollout strategy: simultaneous availability across Kimi.com, Kimi Work, Kimi Code, and the Kimi API.
What This Article Covers
- Kimi K3’s model scale and architecture
- The services and APIs Moonshot AI offers it through
- The plan for releasing model weights and what remains unconfirmed
- What information we deliberately left out of this article, and why
Overview
Moonshot AI announced Kimi K3 on July 16, 2026. According to the official announcement, it has 2.8 trillion total parameters, native image understanding, and supports a context length of up to 1 million tokens. DIGITIMES reported on questions raised within China’s AI industry about Kimi K3, but the publicly available reporting did not let us confirm the specific basis for that criticism. For that reason, this article covers only the product specifications and rollout strategy that we could confirm from Moonshot AI’s own official information.
Details
The specifications confirmed from Moonshot AI’s official announcement are as follows.
- 2.8 trillion total parameters
- Native image understanding
- Context length of up to 1 million tokens
- Architecture using Kimi Delta Attention, Attention Residuals, and Mixture of Experts
- 896 experts, with 16 active per token
- Offered through Kimi.com, Kimi Work, Kimi Code, and the Kimi API
At the time of the announcement, Moonshot AI planned to release the full model weights by July 27, 2026, with a technical report to follow later. Moonshot AI’s own evaluation states that the model does not match top proprietary models in overall performance — note that this assessment comes from the company itself.
Industry Impact
Rolling out a large MoE model simultaneously across multiple services and an API is notable from the standpoint of infrastructure load and compute design needed to support a 1-million-token context in production. The rollout format and compute design deserve more attention here than the model’s raw performance ranking.
DIGITIMES’s headline suggests criticism within China’s AI industry, but the publicly available information did not let us confirm the specific basis for it. Because we do not treat unconfirmed information as fact, this article does not go into the substance of that criticism.
What to Watch Next
- Whether the full model weights and technical report, planned for July 27, 2026, were actually released
- Whether third parties can reproduce Moonshot AI’s own benchmark results
- Training data, training compute, inference cost, and concurrent-usage capacity in production
- The specific basis for the industry criticism DIGITIMES reports
We will update this article as these points become clear.
Summary
Kimi K3 is a Mixture-of-Experts model with 2.8 trillion total parameters, and Moonshot AI has placed simultaneous rollout across multiple services and APIs at the center of its strategy. Whether the model weights were fully released, whether third parties can reproduce the benchmarks, and the training/inference costs all remain to be confirmed. There is also reporting on criticism within China’s AI industry, but this article does not cover claims we could not verify.
