GitHub radar
Kimi K3: MoonshotAI’s Open-Weight Frontier Model
MoonshotAI published the weights of their 2.8 trillion parameter multi-modal agentic model Kimi K3 with 1 million token context size. It’s based on Mixture-of-Experts architecture and supports OpenAI/Anthropic APIs.
Kimi K3 is an open weight model developed by MoonshotAI, a Chinese AI company. This is their main model which is based on Mixture-of-Experts architecture with 2.8 trillion total parameters and 16 active experts per token among 896 tokens. This model has a context size of up to 1 million tokens and can take both images and texts as input using MoonViT-V2 visual encoding. It achieves good scores on various benchmarks like GPQA Diamond 93.5 and MCPMark-Verified 94.5.
Why a vibe-coder should care
You can easily integrate this model into your projects by using it as a drop-in replacement of closed API models. You don’t need to do any changes to your code because this model supports OpenAI/Anthropic APIs. You can self host this model using vLLM/SGLang and get access to a powerful multi modal agentic model without having to pay any additional token fees to third parties.
▌ More finds