Moonshot AI

Kimi K3

Released Jul 16, 2026 · knowledge cutoff unpublished

Status: Frontier
Location: China
Modality: Multimodal
Context window: 1M
Max output: 131K
Speed: —
Price ($/MTok in / out): $3 / $15
Cost per task: $0.72
Open weights: Yes

Benchmarks4 of 8 reported

GPQA Diamond

93.5%

Terminal-Bench 2.1

88.3%

Humanity's Last Exam

43.5%

LMArena Elo

1486

MMLU-Pro

—

SWE-bench Verified

—

AIME (latest)

—

ARC-AGI-2

—

Strengths & weaknesses

Strengths

Agentic browsing and long-horizon coding are where it feels genuinely frontier-class
Native tool-calling conventions that drop into existing agent stacks without adapters
Reads images and long documents natively, unlike earlier Kimi releases

Weaknesses

Slow and very verbose — every request goes through a reasoning pass with no way to switch it off
The effort dial shipped locked at maximum, so cheap quick calls aren't possible
Factual reliability went backwards: it states wrong answers more confidently than K2 did

MMLU-Pro, AIME 2025, and SWE-bench Verified are unconfirmed or conflicting in third-party coverage (AIME cited as both 88% and 96.1%) so left null; Moonshot's own suite reports Terminal-Bench 2.1 at 88.3 and SWE Marathon at 42.0 instead. HLE is 43.5% no tools / 56.0% with tools. Artificial Analysis ranks K3 #3 on its Intelligence Index at 57, with GDPval-AA v2 Elo 1668 (K2.6: 1190), AA-Omniscience +18 on 46% accuracy, and $0.94 per index task at $3.00/$15.00 per MTok — but its hallucination rate regressed to 51% from K2.6's 39%; none of these map to a tracked benchmark key. Weights landed on 2026-07-26 — 96 safetensors shards totalling roughly 1.56 TB, native MXFP4 with MXFP8 activations — under a bespoke Kimi K3 License that attaches conditions to large model-as-a-service and very large commercial deployments, rather than the MIT terms earlier Kimi releases carried.

For developers

API model strings

Not researched

Licence

Not researched

Retirement

No retirement announced

Lineage

No recorded predecessor