Top AI Product

On K3 Open Day, everyone stared at Kimi K3 — the 2.8T-param open model. The quieter drop matters more if you build infra: Kimi AgentENV, the distributed system that actually trains agent models, now open-sourced with kvcache-ai.

What it actually is

AgentENV isn’t a model. It’s the ground floor under one — a platform for running agent environments at massive scale during RL training. It spins up huge numbers of Firecracker microVMs across a cluster, loads any OCI image on demand via overlaybd, and uses local disk as a bounded cache, so your image set can exceed disk without pre-warming every host. Snapshot-backed sandboxes boot or resume in under 50ms and pause under 100ms. A live environment forks into independent copies for parallel rollouts.

Why it’s worth grabbing

Agentic RL needs thousands of concurrent, resettable environments — the orchestration, concurrency, and scheduling headache almost nobody open-sources. Everyone ships agent models; almost no one ships how they trained them. AgentENV is self-hostable: wire it into your own agent RL pipeline and stop rebuilding the environment layer from scratch. If you’re standing up an agent training stack, this fills the real gap.

You Might Also Like

• Ai Agent Book Bojie lis Open Source ai Agent Textbook Gains 1734 Github Stars in one day
• White House Accuses Moonshot ai of Distilling Anthropics Fable to Build Kimi k3 3 4 Million Claude Exchanges one Week Before the Weights Open
• Hermes Agent by Nous Research Might be the Open Source ai Agent That Finally Remembers Everything
• Langchains gtm Agent Drove 250 More Conversions now the Framework Behind it is Open Source
• A Single api String Exposed Cursors Secret Composer 2 Runs on Moonshot ais Kimi k2 5