Overview
moonshot-ai’s July 16, 2026 frontier MoE model: 2.8 trillion total parameters, native vision, 1M-token context. Marketed as the largest open-source/open-weight class model; full weights scheduled July 27, 2026 (Modified MIT). Until then, Artificial Analysis treats the live API as proprietary hosted.
Specs (vendor-reported)
- MoE: 16 of 896 experts activated (2026-07-17-moonshot-kimi-k3-marktechpost)
- Architecture: kimi-delta-attention (KDA) + AttnRes
- API pricing commonly cited ~15 per M tokens; OpenAI-SDK compatible; OpenRouter id
moonshotai/kimi-k3 - Surfaces: kimi.com, Kimi Work, Kimi Code CLI, API
Benchmarks
- Self + artificial-analysis: near Claude Fable 5 / GPT-5.6 Sol on select agentic/coding; trails top proprietary overall
- MarkTechPost: leads BrowseComp, SWE Marathon, OmniDocBench; trails Fable 5 on FrontierSWE / HLE-Full
Weights not public at launch — "open-source" framing is prospective until Jul 27 checkpoint.
Recent Developments
-
2026-10-22: cerebras used Kimi K3 (~3T params, 69 linear + 24 dense layers) as illustrative wafer-scale inference example—59 WSE-3T chips (2026-10-25-cerebras-trillion-parameter-frontier-models)
-
2026-09-23: fireworks-ai released ember-1 — K3 post-trained for ~40% fewer reasoning tokens; SWE-bench 92.2%, Terminal Bench 82.0% (vendor-reported) (2026-09-27-fireworks-ember-1-efficient-reasoning)
-
2026-07-19: New consumer subscriptions paused after 48h demand surge; existing users prioritized; Code membership split (moonshot-kimi-k3-subscription-pause)
-
2026-07-20: IPO path news amplifies K3 demand narrative (moonshot-ai-hong-kong-ipo-30b)
-
2026-07-19: Scale context for qwen-38 (2.4T) — K3 remains largest publicly disclosed Chinese model at 2.8T (2026-07-19-alibaba-qwen-38-scmp)
-
2026-07-16: API launch ahead of WAIC Shanghai (2026-07-17-moonshot-kimi-k3-28t-open-source)
-
2026-07-27: Planned full weight release