Kimi K3

Moonshot AI·2026-07-16

LLMsopen-weightcloud

Overview

2.8T MoE (~50B active, 16 of 896 experts) native multimodal flagship with a 1M-token context window. Launched via API, app and playground on 2026-07-16; open weights scheduled for 2026-07-27.

Capabilities and innovations

2.8T total / ~50B active MoE (896 experts, 16 routed)1M token context windowNatively multimodal (text, image, video)Vision-in-the-loop (screenshot inspect + code edit)Open weights scheduled 2026-07-27Kimi Delta Attention2.8T Open-Weight MoE (largest open-weight at launch)Vision-in-the-Loop Agent Feedback1M-Token Native Context

Benchmarks

Benchmark	Score	Source
GPQA Diamond	93.5%	vendor GPQA-Diamond 93.5 (official launch blog, 2026-07-16).
Humanity’s Last Exam (no tools)	43.5%	vendor HLE-Full 43.5 no-tools (text-only), recorded for cross-model comparability. Moonshot also reports 56.0 with tools, which is not comparable to the no-tools HLE used here.
MMLU-Pro	87.97%	independent MMLU Pro 87.97%, measured by Vals AI (accessed 2026-08-17). Moonshot publishes no MMLU-Pro value. Vals AI measures GPQA Diamond at 92.93 on the same run, 0.6 points below Moonshot's self-reported 93.5, so this harness reads in line with the vendor table.
Terminal-Bench 2.x	88.3%	vendor Terminal-Bench 2.1 88.3 (official launch blog).
MMMU-Pro	81.6%	vendor MMMU-Pro 81.6 (official launch blog).

SWE-bench Verified: not reported by the vendor
SWE-bench Pro: not reported by the vendor — Moonshot reports its own coding suite (Program Bench, SWE Marathon, FrontierSWE, DeepSWE) instead of SWE-bench Pro.

API pricing

$3 per million input tokens, $15 per million output tokens (USD, cheapest OpenRouter endpoint, checked 2026-07-17)

Architecture and hardware

Parameters: 2800B total, 50B active per token (MoE)

Reliability

tau3_banking: 46 ()third party

More from Moonshot AI

Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.