MarkTechPost (Jul 16, 2026): Technical summary of Moonshot Kimi K3.
- 2.8T-parameter MoE; activates 16 of 896 experts; native vision; 1M context
- Architecture: Kimi Delta Attention (KDA), Attention Residuals (AttnRes); ~2.5x better scaling vs K2 claimed
- Live on Kimi.com, Kimi Work, Kimi Code, API (OpenAI SDK compatible); model id kimi-k3
- Pricing cited: 3.00/$15.00 per MTok tiers (input/output/reasoning variants in coverage)
- Benchmarks: leads BrowseComp, SWE Marathon, OmniDocBench; trails Claude Fable 5 on FrontierSWE and HLE-Full
- Company calls it first open 3T-class model; full weights still pending
Pricing tiers differ across outlets ( 15 common; MarkTechPost lists additional tier) — treat as vendor-reported.