Skip to main content

Moonshot AI

Kimi K3

Open-weight frontier model scoring within a few points of the closed leaders on agentic coding.

Catalog record checked August 4, 2026; individual provider fields may change.

AI model specification details
SpecificationKimi K3
ProviderMoonshot AI
TierFrontier
Context window1.05M
Max output1M
Input / 1M tokens$3
Output / 1M tokens$15
WeightsOpen
Parameters2.8T total / 104B active (MoE)
Modalitiestext, image, video
ReleasedJuly 16, 2026

Verified evidence

Published benchmark results

Each result keeps its source and measurement date visible. A missing benchmark is not treated as a zero.

Artificial Analysis Intelligence Index [max]

2026-08-08 · Artificial Analysis

Benchmark

DeepSWE 1.1 in context

The full local snapshot puts this model beside the wider field, including cost per completed task.

Local leader

Claude Opus 5 [max]

74%

Rows shown

24

Highest published reasoning effort per model (not best Pass@1)

Snapshot date

2026-08-13

Mirrored from deepswe.datacurve.ai

DeepSWE 1.1 pass@10%16%32%48%64%80%$0$4.50$9.00$13.50$18.00$22.50$27.00Avg cost per taskClaude Opus 5 [max]: 74% · $11.84 · 118k tokens · 99 stepsClaude Opus 5GPT-5.6 Sol [max]: 73% · $8.39 · 60k tokens · 61 stepsGPT-5.6 SolClaude Fable 5 [max]: 70% · $21.63 · 119k tokens · 88 stepsClaude Fable 5GPT-5.6 Terra [max]: 70% · $4.95 · 72k tokens · 76 stepsGPT-5.6 TerraKimi K3 [max]: 69% · $4.65 · 82k tokens · 98 stepsKimi K3GPT-5.6 Luna [max]: 67% · $3.03 · 73k tokens · 102 stepsGPT-5.6 LunaGPT-5.5 [xhigh]: 67% · $7.23 · 46k tokens · 82 stepsGPT-5.5Grok 4.6 [xhigh]: 67% · $5.50 · 71k tokens · 87 stepsGrok 4.6Gemini 3.7 Flash [high]: 65% · $2.18 · 107k tokens · 125 stepsGemini 3.7 FlashDeepSeek V4-Pro [max]: 63% · $0.24 · 106k tokens · 155 stepsDeepSeek V4-ProClaude Opus 4.8 [max]: 59% · $13.22 · 135k tokens · 120 stepsClaude Opus 4.8Qwen3.8-Max [xhigh]: 58% · $3.73 · 95k tokens · 111 stepsQwen3.8-MaxMuse Spark 1.2 [xhigh]: 55% · $3.70 · 99k tokens · 101 stepsMuse Spark 1.2Claude Sonnet 5 [max]: 54% · $26.40 · 214k tokens · 268 stepsClaude Sonnet 5Grok 4.5 [high]: 54% · $2.42 · 36k tokens · 61 stepsGrok 4.5DeepSeek V4-Flash [max]: 53% · $0.10 · 108k tokens · 153 stepsDeepSeek V4-FlashMuse Spark 1.1 [xhigh]: 53% · $2.36 · 74k tokens · 96 stepsMuse Spark 1.1GPT-5.4 [xhigh]: 52% · $5.65 · 71k tokens · 70 stepsGPT-5.4Gemini 3.6 Flash [high]: 47% · $4.42 · 96k tokens · 117 stepsGemini 3.6 FlashGLM 5.2 [max]: 44% · $3.92 · 78k tokens · 129 stepsGLM 5.2Gemini 3.5 Flash [high]: 36% · $3.45 · 76k tokens · 105 stepsGemini 3.5 FlashKimi K2.7 Code: 31% · $2.82 · 59k tokens · 149 stepsKimi K2.7 CodeClaude Sonnet 4.6 [high]: 30% · $5.52 · 76k tokens · 134 stepsClaude Sonnet 4.6Gemini 3.1 Pro [high]: 12% · $2.14 · 28k tokens · 76 stepsGemini 3.1 Pro

Better is toward the top-right (higher pass rate, lower cost). X-axis is reversed to match DeepSWE’s public chart. v1.1 uses average cost / tokens / steps; v1 uses published medians.

DeepSWE 1.1 leaderboard with pass rate, cost, tokens, and steps per task
#ModelPass@1Cost / taskTokens / taskSteps / task
1Claude Opus 5 [max]74%$11.84118k99
2GPT-5.6 Sol [max]73%$8.3960k61
3Claude Fable 5 [max]70%$21.63119k88
4GPT-5.6 Terra [max]70%$4.9572k76
5Kimi K3 [max]69%$4.6582k98
6GPT-5.6 Luna [max]67%$3.0373k102
7GPT-5.5 [xhigh]67%$7.2346k82
8Grok 4.6 [xhigh]67%$5.5071k87
9Gemini 3.7 Flash [high]65%$2.18107k125
10DeepSeek V4-Pro [max]63%$0.24106k155
11Claude Opus 4.8 [max]59%$13.22135k120
12Qwen3.8-Max [xhigh]58%$3.7395k111
13Muse Spark 1.2 [xhigh]55%$3.7099k101
14Claude Sonnet 5 [max]54%$26.40214k268
15Grok 4.5 [high]54%$2.4236k61
16DeepSeek V4-Flash [max]53%$0.10108k153
17Muse Spark 1.1 [xhigh]53%$2.3674k96
18GPT-5.4 [xhigh]52%$5.6571k70
19Gemini 3.6 Flash [high]47%$4.4296k117
20GLM 5.2 [max]44%$3.9278k129
21Gemini 3.5 Flash [high]36%$3.4576k105
22Kimi K2.7 Code31%$2.8259k149
23Claude Sonnet 4.6 [high]30%$5.5276k134
24Gemini 3.1 Pro [high]12%$2.1428k76

DeepSWE “Best” picks the highest published reasoning effort per model (not the highest pass rate). Small gaps may not be statistically meaningful — confirm on deepswe.datacurve.ai.

Best for

  • Agentic coding without vendor lock-in
  • Self-hosting at frontier quality
  • Long-context work

Watch out

2.8T parameters means self-hosting is a datacentre exercise, not a workstation one — open weights here mean provider choice, not local inference.

Common questions

Kimi K3

Answered from the verified figures on this page rather than general guidance.

How much does Kimi K3 cost per million tokens?

Kimi K3 is listed at $3 per million input tokens and $15 per million output tokens at standard rates. Output tokens usually dominate real bills, so weigh the output rate more heavily than the input rate.

What is Kimi K3's context window?

Kimi K3 accepts about 1.05M tokens of context. That only matters if you routinely send very long documents, large codebases, or multi-turn histories that approach that limit.

What is Kimi K3 best for?

Kimi K3 is a frontier tier from Moonshot AI. It suits agentic coding without vendor lock-in, self-hosting at frontier quality, long-context work. 2.8T parameters means self-hosting is a datacentre exercise, not a workstation one — open weights here mean provider choice, not local inference.

Can I self-host Kimi K3?

Kimi K3 publishes open weights, but self-hosting depends on the licence, hardware footprint, quantisation quality, and serving stack. A hosted API is often cheaper until you have measured throughput and concurrency on your own hardware.

Compare it

Head-to-head model comparisons

These are the published pairings that put this model against a plausible alternative.