Kimi K3 launches without a model card — 1M context is limited to Allegretto and above
TL;DR
Kimi K3 is rolling out: Moderato supports 256K context, while Allegretto and above reach up to 1M.
Moonshot AI abruptly began rolling the Kimi K3 AI model out to users on the evening of July 16, raising the maximum context window from 256K to 1M tokens. Kimi's website now leads with K3, promoting coding, 3D game creation, slides, and parallel task execution through Swarm and Goal.
1M is not the default allowance for every user. Kimi Code offers K3 and K2.7 Code across three model IDs. Moderato has a 256K context window, Allegretto and higher tiers reach up to 1M, and Andante is not yet available.
K3 reached the product before its technical documentation. When IT Home published, Kimi's website was already showing K3, but its official blog had no release announcement. Moonshot had also not published a model card, parameter count, benchmarks, API pricing, or complete limits for each tier.
Kimi released the MoE-based K2 in July 2025 with 1T total parameters and 32B activated per inference. The open-source K2.6 released in April 2026 expanded Agent Swarm to 300 sub-agents. K2.7 Code followed in June with 30% lower reasoning-token use and remains alongside K3.
The window has expanded fourfold from 256K to 1M. Whether an account can use the full 1M depends first on whether it receives Moderato or Allegretto and above.
via Kimi / IT Home / DoNews
1M is not the default allowance for every user. Kimi Code offers K3 and K2.7 Code across three model IDs. Moderato has a 256K context window, Allegretto and higher tiers reach up to 1M, and Andante is not yet available.
K3 reached the product before its technical documentation. When IT Home published, Kimi's website was already showing K3, but its official blog had no release announcement. Moonshot had also not published a model card, parameter count, benchmarks, API pricing, or complete limits for each tier.
Kimi released the MoE-based K2 in July 2025 with 1T total parameters and 32B activated per inference. The open-source K2.6 released in April 2026 expanded Agent Swarm to 300 sub-agents. K2.7 Code followed in June with 30% lower reasoning-token use and remains alongside K3.
The window has expanded fourfold from 256K to 1M. Whether an account can use the full 1M depends first on whether it receives Moderato or Allegretto and above.
via Kimi / IT Home / DoNews
