Moonshot AI Releases Kimi K3, First Open 3T-Class Model
The 2.8-trillion-parameter model with 1M-token context is now available on Hugging Face under a Modified MIT license.
Moonshot AI has released Kimi K3, a 2.8-trillion-parameter open-weight model that the Beijing-based startup describes as the world's first open 3T-class system. The model weights went live on Hugging Face on July 27, 2026, under a Modified MIT license, following an API launch on July 16.
Architecture and Scale
Kimi K3 uses a highly sparse Mixture-of-Experts architecture with 896 experts, activating only 16 per token. The model introduces two new components: Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), designed to stabilize training and inference at this scale.
The system supports a 1-million-token context window and processes images natively alongside text. Moonshot positions K3 for long-horizon coding, knowledge work, and reasoning tasks that require sustained attention across massive documents or codebases.
Open Weights, Open Questions
The Hugging Face release (moonshotai/Kimi-K3) gives researchers and developers direct access to a model at a scale previously reserved for closed systems. This could accelerate work on autonomous agents and complex multi-step workflows in the open ecosystem.
Moonshot's self-reported benchmarks claim a GPQA Diamond score of 93.5%, though this figure has not been independently verified by academic evaluators. Third-party trackers including aireleasetracker.com have published the number, but it should be treated as the company's own assessment rather than a peer-reviewed result.
Safety Concerns
A preliminary assessment by the UK's AISI and CAISI found that Kimi K3 demonstrated limited autonomous cyberattack capability when given network access. The model solved one of ten tasks in the TLO cyber range and failed to achieve arbitrary code execution on ExploitBench (0 of 41 samples). The agencies noted the capability exists but trails frontier US models significantly on cybersecurity benchmarks.
What Comes Next
The Modified MIT license governing Kimi K3 is more permissive than many frontier model releases, though researchers should review the specific terms before deployment.
For the open-source community, Kimi K3 represents a threshold moment: a model at genuine frontier scale, available for inspection, fine-tuning, and integration without the gates that typically surround 3T-class systems. Whether it delivers on Moonshot's intelligence claims in real-world workloads remains to be seen.