Overview
- Moonshot completed its public rollout by publishing full Kimi K3 weights, a technical report and supporting code to Hugging Face and GitHub on Monday, July 27, making the model and infrastructure downloadable under a Kimi K3 License.
- Kimi K3 is a sparse Mixture‑of‑Experts design with 2.8 trillion total parameters, about 104 billion activated per token through 896 experts selecting 16 per token, and it supports a one‑million‑token context window and native vision input.
- Independent leaderboards and community benchmarks placed Kimi K3 among the top models for agentic and coding tasks, with Agent Arena showing K3 as the highest‑performing open‑weight model and in the top three overall.
- Running K3 locally is resource intensive because the weights are roughly 1.4 TB and practical self‑hosting requires large distributed GPU clusters, yet commercial hosts such as Telnyx have begun offering hosted inference with published pricing.
- Anthropic and U.S. officials have accused Moonshot of using distillation from U.S. frontier models during development, Moonshot has denied the claim and pointed to architectural innovations, and the dispute remains unresolved with potential procurement and regulatory implications for buyers.