Simon Willison's Weblog·2 months ago·
46
● 33 sources
Moonshot AI released the weights for Kimi K3, a 2.8 trillion parameter model, totaling 1.56TB on Hugging Face. The model's license requires a separate commercial agreement with Moonshot for any Model as a Service business exceeding $20 million in annual revenue. Kimi avoids calling this "open source," instead using "open weight," and OpenRouter is already offering K3 from multiple providers at comparable pricing to Moonshot's official rates.
Moonshot AI's Kimi team and kvcache-ai open-sourced AgentENV, a distributed platform for running isolated Linux environments needed to train language models with reinforcement learning on real computer tasks. The system uses Firecracker microVMs with shared read-only storage layers and memory ballooning to run training at scale, with each sandbox maintaining its own kernel and filesystem while reducing startup overhead. This infrastructure enables agentic RL training for Kimi K3, a 2.8-trillion-parameter model, making it feasible to train agents that interact with real computing systems.
Moonshot AI released open weights for Kimi K3, a 2.8-trillion-parameter language model, on Hugging Face following overwhelming API demand. Running the model requires approximately 1.4 TB of storage and a distributed GPU environment with at least eight servers equipped with eight NVIDIA H100 or B200 accelerators each. Organizations can now self-host K3 to avoid recurring API costs, though the substantial infrastructure investment makes this option practical primarily for those with strict regulatory requirements or existing GPU capacity.
Moonshot AI, a Chinese startup, released Kimi K3, an AI model that performs comparably to leading US systems at lower cost, and plans to distribute the model's weights for free to developers worldwide. The model reportedly matches or exceeds performance of some top US AI systems while being significantly cheaper to operate. This shift toward open-weight models threatens the market dominance of closed proprietary systems from US companies by giving developers greater control and accessibility to capable alternatives.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
