Submit Tool

Kimi K3

Moonshot's 2.8 trillion parameter open-weight model, the largest open model shipped to date

Free Multimodal
Visit Product Page →
1000000 context tokens
2.8T (MoE) parameters
Modified MIT license
Jul 2026 released

Kimi K3 is Moonshot AI’s largest release yet, a 2.8 trillion parameter mixture-of-experts model that the company says is the biggest open-weight model shipped by anyone so far. It reads text, images, and video natively, carries a 1 million token context window, and runs in an always-on thinking mode rather than offering a separate fast path. Moonshot priced API access at $3 per million input tokens (dropping to $0.30 on a cache hit) and $15 per million output tokens, undercutting closed frontier models from Western labs by a wide margin. Full weights landed on Hugging Face shortly after the July 16 announcement, using MXFP4 quantization to bring the download to roughly 1.4 TB, small enough for serious self-hosting setups to run without a dedicated cluster.

// similar models in Multimodal

Llama 5 Meta's next flagship open-weight multimodal model, built for on-device and datacenter use alike Meta AI · Multimodal · Free
Gemini 3 Pro Google's flagship multimodal model with a 1M-token context window Google DeepMind · Multimodal · Paid
Gemini 3.5 Pro Google's next-generation frontier model, announced but not yet fully released Google DeepMind · Multimodal · Paid
GPT-4o OpenAI's omni model, natively multimodal across text, image, and audio OpenAI · Multimodal · Paid
Gemini 3.7 Flash Google's successor to Gemini 3.6 Flash, shipped three weeks later with cheaper output pricing Google DeepMind · Multimodal · Paid
Gemini 3.6 Flash Google's new default Flash model, cheaper and faster on coding and agentic tasks than 3.5 Flash Google DeepMind · Multimodal · Paid