On July 26, 2026, Moonshot AI released the full weights for Kimi K3 on Hugging Face — 96 shards totalling approximately 1.56 TB. With 2.8 trillion total parameters and 104 billion active per token, it is now the largest open-weight model ever made publicly available.
The model is a Mixture-of-Experts architecture. At inference time it only activates a fraction of its parameter count, which is how it remains somewhat practical to self-host despite its enormous footprint. Day-0 hosting was available through Together AI and Modal within hours of the weights going live, so you don't necessarily need to own a server rack to use it — but the option is now there.
What it can do
Kimi K3 accepts text, images, and video, and supports a 1-million-token context window. In arena blind evaluations, developers selected it over leading U.S. models on front-end coding tasks. It benchmarks below Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol in overall rankings, but outperforms prior-generation flagship models from both companies across coding and general agent evaluations.
The licence catch
The weights are out there, but they're not MIT-licensed. Moonshot AI released them under a custom commercial licence — a distinction that matters for enterprise deployments. Before building anything production-facing on Kimi K3, read the terms carefully. The research and experimentation community tends to move first and worry about licences later, but compliance teams take a different view.
Why self-hosting matters here
The practical significance of open weights is often undersold in the headline numbers. A 2.8-trillion-parameter model is impressive, but what actually changes when weights are public is the risk profile for data-sensitive deployments. When you self-host, your prompts and completions never cross a Chinese API endpoint — a concern that had been raised repeatedly about Kimi K3's API version since the model's launch earlier this year. This is what Nathan Lambert described as "the open-weights escalation": the race for frontier-grade open models is compressing the gap between what you can download and what only hyperscalers used to run.
For developers and researchers comfortable managing 1.56 TB of model weights, the question is no longer whether open-source models can compete — it's how far ahead proprietary labs can stay.