Kimi K2 models

Strongest: Kimi K2.7-Code

Kimi K2.7-Code is Moonshot AI's coding-specialized trillion-parameter MoE (32B active, 384 experts, MLA attention), released 2026-06-12 as open weights under a modified-MIT license with a 256K context, always-on thinking preserved across turns, and native multimodal input including video. It claims roughly 30% fewer thinking tokens than K2.6 at $0.95/$4.00 per 1M tokens — but its headline benchmark numbers are vendor-run on Moonshot's own suites, and practitioners have publicly disputed that they hold up; independent third-party coding scores land well below the closed frontier. Read it as a cheap, genuinely capable open coding agent with an unsettled evidence base, not a verified frontier contender.