Patchwire

Moonshot AI · Released Jul 16, 2026

Kimi K3

Kimi K3 is a 2.8T parameter open-weight multimodal reasoning model from Moonshot AI. It is suited for complex coding, knowledge work, and long-horizon agentic workflows, and is particularly strong at...

ReasoningOpen weightsVision
  • Top 10%SciCode
  • Top 11%WebDev Arena
  • Top 15%ProofBench
  • Top 19%GPQA Diamond
Compare in showdown → OpenRouter ↗ All Moonshot AI models
Released
Jul 16 2026
Context window
1M 944K out
Input / 1M tokens
$0.24 from $0.24
Output / 1M tokens
$11.94 from $3.00

Where to run it

20 providers compared

Cheapest: Relace at $0.80 / $3.00 per 1M input / output tokens — 85% less than the priciest option.

ProviderInputOutputCache readContextMax outputUptimevs. best
RelaceFP4 $0.80 $3.00 $0.601M944K98.8%24h Cheapest
Sail ResearchFP4 $1.20 $9.00 $0.401M944K100%24h +133%
Sail Researchus · FP4 · no tool calling $1.20 $9.00 $0.401M944K99.9%24h +133%
InferenceNetFP4 $0.24 $11.94 $0.211M944K99.5%24h +134%
Waferus · MXFP4 $0.36 $11.90 $0.361M944K100%24h +140%
Decartmxfp4 · MXFP4 $1.65 $8.25 $0.171M944K99%24h +144%
Chutesmxfp4 · MXFP4 · no tool calling $1.85 $9.50 $0.181M66K89.5%24h +179%
WaferMXFP4 $0.24 $16.00 $0.231M944K100%24h +210%
Fireworksno tool calling $2.10 $10.50 $0.211M944K0%24h +211%
MakoraFP4 $0.65 $15.00 $0.231M944K99.4%24h +214%
AkashMLFP4 $1.20 $14.00 $1.201M944K100%24h +226%
MorphFP8 $0.91 $14.90 $0.441M944K99.8%24h +227%
Phala $2.55 $12.75 $0.261M944K99.8%24h +278%
DigitalOcean $2.55 $12.95 $0.261M944K99.8%24h +281%
InferenceNetfast · FP4 $2.24 $14.52 $0.351M944K99.4%24h +293%
DeepInframxfp4 · MXFP4 $2.85 $14.25 $0.281M944K99.8%24h +322%
Modalmxfp4 · MXFP4 $3.00 $15.00 $0.301M944K100%24h +344%
BaseTenFP8 $3.00 $15.00 $0.301M262K100%24h +344%
Moonshot AImxfp4 · MXFP4 $3.00 $15.00 $0.301M944K100%24h +344%
Fireworks $3.00 $15.00 $0.301M944K99.9%24h +344%
Together $3.00 $15.00 $0.301M944K99.6%24h +344%
Parasailfast · FP4 · no tool calling $3.00 $15.00 $0.301M900K95.3%24h +344%
ParasailFP4 $3.00 $15.00 $0.301M944K91.5%24h +344%
Amazon Bedrockus-east-2 $3.00 $15.00 $0.301M128K85.5%24h +344%
Alibaba $3.45 $17.25 $0.341M131K98.8%24h +411%
Fireworksus $4.50 $22.50 $0.451M944K100%24h +567%
Fireworksfast $4.50 $22.50 $0.451M944K98.9%24h +567%

Value

Price vs. performance

On the value frontierNo cheaper model beats Kimi K3 on GPQA Diamond.
25.3%49.0%72.7%96.4% $0.01$0.10$1.00$10.00$100.00 DeepSeek V4.1 Flash 1017 (DeepSeek) — 89.8% · $0.30/1M in Gemini 4 Argon (Google) — 96.4% · $2.00/1M in Claude Haiku 5.5 (Anthropic) — 89.6% · $0.10/1M in Mistral Large 4 (Mistral) — 88.1% · $0.68/1M in GPT-6.1 Sol (OpenAI) — 95.4% · $2.00/1M in Claude Sonnet 5.5 (Anthropic) — 95.6% · $2.00/1M in GPT-6 Luna (OpenAI) — 90.5% · $0.10/1M in GPT-6 Sol (OpenAI) — 94.3% · $2.00/1M in Claude Opus 5.5 (Anthropic) — 90.6% · $4.00/1M in Grok 4.7 (xAI) — 92.7% · $2.00/1M in DeepSeek V4.1 Flash (DeepSeek) — 89.8% · $0.056/1M in GPT-6 Astra (OpenAI) — 95.8% · $10.00/1M in Qwen3.8 Max (0902) (Qwen) — 92.7% · $2.00/1M in Gemini 3.8 Flash (Google) — 95.4% · $0.82/1M in GLM 5.3 Flash (Z.ai) — 90.1% · $0.047/1M in GLM 5.3 (Z.ai) — 90.9% · $0.079/1M in Gemini 3.7 Flash (Google) — 94.8% · $0.75/1M in DeepSeek V4 Pro 0813 (DeepSeek) — 91.7% · $0.66/1M in Grok 4.6 (xAI) — 94.0% · $2.00/1M in DeepSeek V4 Flash 0731 (DeepSeek) — 91.0% · $0.44/1M in Qwen3.7 Flash (Qwen) — 82.3% · $0.03/1M in Claude Opus 5 (Anthropic) — 93.9% · $5.00/1M in Gemini 3.6 Flash (Google) — 94.1% · $0.75/1M in Gemini 3.5 Flash Lite (Google) — 83.3% · $0.30/1M in GPT-5.6 Luna (OpenAI) — 91.6% · $0.20/1M in GPT-5.6 Terra (OpenAI) — 93.3% · $2.00/1M in GPT-5.6 Sol (OpenAI) — 93.5% · $4.00/1M in Grok 4.5 (xAI) — 93.4% · $2.00/1M in Claude Sonnet 5 (Anthropic) — 90.5% · $2.00/1M in GLM 5.2 (Z.ai) — 91.9% · $0.62/1M in Claude Fable 5 (Anthropic) — 85.9% · $10.00/1M in Qwen3.7 Plus (Qwen) — 87.9% · $0.32/1M in MiniMax M3 (MiniMax) — 90.9% · $0.30/1M in Claude Opus 4.8 (Anthropic) — 91.0% · $5.00/1M in Qwen3.7 Max (Qwen) — 90.9% · $1.48/1M in Gemini 3.5 Flash (Google) — 92.8% · $1.50/1M in Gemini 3.1 Flash Lite (Google) — 81.8% · $0.25/1M in Grok 4.3 (xAI) — 88.8% · $1.25/1M in Qwen3.5 Plus 2026-04-20 (Qwen) — 84.9% · $0.30/1M in Qwen3.6 Flash (Qwen) — 83.3% · $0.19/1M in Qwen3.6 35B A3B (Qwen) — 84.9% · $0.10/1M in Qwen3.6 27B (Qwen) — 85.9% · $0.25/1M in GPT-5.5 (OpenAI) — 90.7% · $5.00/1M in DeepSeek V4 Pro 0423 (DeepSeek) — 91.7% · $0.95/1M in DeepSeek V4 Flash 0423 (DeepSeek) — 91.0% · $0.091/1M in Claude Opus 4.7 (Anthropic) — 90.1% · $5.00/1M in GLM 5.1 (Z.ai) — 89.9% · $0.97/1M in Gemma 4 26B A4B (Google) — 73.2% · $0.09/1M in Gemma 4 31B (Google) — 75.8% · $0.20/1M in Qwen3.6 Plus (Qwen) — 88.4% · $0.33/1M in GPT-5.4 Nano (OpenAI) — 78.5% · $0.20/1M in GPT-5.4 Mini (OpenAI) — 86.9% · $0.75/1M in Mistral Small 4 (Mistral) — 47.5% · $0.15/1M in Qwen3.5-9B (Qwen) — 79.0% · $0.10/1M in GPT-5.4 Pro (OpenAI) — 94.6% · $30.00/1M in GPT-5.4 (OpenAI) — 93.3% · $2.50/1M in Gemini 3.1 Flash Lite Preview (Google) — 81.8% · $0.25/1M in Qwen3.5-35B-A3B (Qwen) — 83.5% · $0.15/1M in Gemini 3.1 Pro Preview (Google) — 94.4% · $2.00/1M in Claude Sonnet 4.6 (Anthropic) — 87.4% · $3.00/1M in Qwen3.5 397B A17B (Qwen) — 86.4% · $0.39/1M in GLM 5 (Z.ai) — 87.8% · $0.60/1M in Claude Opus 4.6 (Anthropic) — 90.5% · $5.00/1M in GLM 4.7 Flash (Z.ai) — 60.5% · $0.061/1M in GLM 4.7 (Z.ai) — 83.3% · $0.54/1M in Gemini 3 Flash Preview (Google) — 89.4% · $0.50/1M in GPT-5.2 (OpenAI) — 91.4% · $1.75/1M in Ministral 3 8B 2512 (Mistral) — 27.2% · $0.15/1M in Ministral 3 3B 2512 (Mistral) — 25.3% · $0.10/1M in Mistral Large 3 2512 (Mistral) — 51.3% · $0.50/1M in Claude Opus 4.5 (Anthropic) — 86.1% · $5.00/1M in GPT-5.1 (OpenAI) — 87.6% · $1.25/1M in Claude Haiku 4.5 (Anthropic) — 71.2% · $1.00/1M in Claude Sonnet 4.5 (Anthropic) — 82.3% · $3.00/1M in GPT-5 (OpenAI) — 86.2% · $1.25/1M in GPT-5 Mini (OpenAI) — 75.0% · $0.25/1M in GPT-5 Nano (OpenAI) — 69.4% · $0.05/1M in gpt-oss-120b (OpenAI) — 75.8% · $0.15/1M in gpt-oss-20b (OpenAI) — 60.8% · $0.018/1M in Claude Opus 4.1 (Anthropic) — 77.3% · $15.00/1M in Qwen3 30B A3B Instruct 2507 (Qwen) — 55.6% · $0.09/1M in Qwen3 235B A22B Thinking 2507 (Qwen) — 80.0% · $0.45/1M in Qwen3 235B A22B Instruct 2507 (Qwen) — 70.7% · $0.09/1M in Gemini 2.5 Pro (Google) — 85.3% · $1.25/1M in Gemini 2.5 Pro Preview 06-05 (Google) — 85.3% · $1.25/1M in R1 0528 (DeepSeek) — 76.3% · $0.50/1M in Qwen3 30B A3B (Qwen) — 61.7% · $0.12/1M in Qwen3 14B (Qwen) — 63.8% · $0.12/1M in Qwen3 32B (Qwen) — 65.7% · $0.08/1M in Kimi K2.7 Code (Moonshot AI) — 87.9% · $0.67/1M in Kimi K2.6 (Moonshot AI) — 90.8% · $0.65/1M in Kimi K3 (Moonshot AI) — 93.1% · $0.24/1M in Kimi K3

Kimi K3 Other Moonshot AI Other labs Value frontier · GPQA Diamond vs. input price per 1M tokens (log scale) · Drag to pan, pinch or Ctrl + scroll to zoom, double-click to zoom in.

Benchmarks

25 benchmark scores

BenchmarkScoreRankLeader
FrontierSWESoftware engineering · Ultra-long-horizon engineering tasks with a 20-hour budget each 25.9% #12of 19 GPT-6 Astra65.5%
DeepSWESoftware engineering · Original long-horizon tasks across active open-source repos 68.5% #8of 26 GPT-6 Astra74.1%
FrontierCodeSoftware engineering · Mergeable fixes for hard open-source issues 44.2% #15of 41 Claude Opus 5.554.6%
GBAEvalSoftware engineering · Build a Game Boy Advance emulator from scratch 48.4% #9of 23 Claude Opus 579.6%
SciCodeSoftware engineering · Research-coding problems from the natural sciences 59.5% #9of 93 Claude Opus 5.566.9%
WeirdML v3Software engineering · Unfamiliar ML and data tasks with vague goals 7.5% #14of 22 GPT-6 Astra42.2%
APEX-AgentsAgents & computer use · Long professional tasks in banking, consulting and law 47.8% #28of 51 Gemini 4 Argon82.4%
Vending-Bench 2Agents & computer use · Run a simulated vending business for a year (final balance) $5,165 #27of 61 GPT-6 Astra$15,515
PostTrainBenchAgents & computer use · Post-train small models under a fixed compute budget 32.0% #5of 11 Claude Fable 541.8%
GDP.pdfWriting & content · Answer questions about real professional PDFs 19.0% #24of 38 GPT-6.1 Sol35.0%
SimpleQA VerifiedWriting & content · Factual accuracy on 1,000 short questions 50.6% #25of 66 Gemini 4 Argon76.0%
WebDev ArenaDesign & visual · People vote on which model builds the better website (Elo) 1674 #9of 88 Claude Opus 5.51820
Blueprint-Bench 2Design & visual · Turn apartment photos into 2D floor plans 29.5% #19of 32 Gemini 4 Argon54.4%
Furniture AssemblyDesign & visual · Spot mistakes in photos of IKEA builds 34.2% #20of 34 Claude Opus 5.583.3%
GPQA DiamondReasoning & knowledge · PhD-level biology, chemistry and physics questions 93.1% #17of 92 Gemini 4 Argon96.4%
ARC-AGI-2Reasoning & knowledge · Few-shot abstract reasoning on grid puzzles 60.4% #33of 63 GPT-6 Astra95.0%
ARC-AGI-1Reasoning & knowledge · Pattern induction from input–output examples 94.5% #16of 63 Gemini 4 Argon98.5%
SimpleBenchReasoning & knowledge · Common-sense trick questions about space, time and people 60.7% #21of 43 Claude Fable 581.9%
CritPtReasoning & knowledge · Open-ended, research-level physics problems 23.4% #18of 96 GPT-5.6 Sol32.3%
Chess PuzzlesReasoning & knowledge · Find the single best move in 100 new puzzles 37.0% #38of 70 GPT-6 Astra76.0%
LMCAReasoning & knowledge · Judge the quality of conceptual arguments like experts 52.7% #20of 96 Claude Sonnet 5.568.3%
FrontierMathMaths · Expert-written problems, undergrad to early research (tiers 1–3) 72.2% #23of 72 GPT-6.1 Sol93.7%
FrontierMath Tier 4Maths · Exceptionally hard research-level problems 39.0% #23of 62 GPT-6.1 Sol100.0%
ProofBenchMaths · Graduate maths proofs verified in Lean 4 87.0% #10of 70 Claude Sonnet 5.5100.0%
OTIS Mock AIMEMaths · Competition-style problems harder than MATH Level 5 97.2% #29of 90 Gemini 4 Argon100.0%

Specs

Details

Model ID
moonshotai/kimi-k3
Developer
Moonshot AI
Release date
Jul 16, 2026 ·
Context window
1,048,576 tokens
Max output
943,718 tokens
API price
$0.24 / $11.94 per 1M input / output tokens
Providers
20
Input types
text, image, video
Reasoning
Yes
Tool calling
Yes
Open weights
Yes
Free tier
No

Release date, prices and providers from OpenRouter's catalog; buying directly from a provider can cost a little differently. Uptime is the share of OpenRouter requests each provider served over the last day. Benchmark scores are the best result per benchmark from Epoch AI, licensed CC BY 4.0. Rank counts every model tracked here with a score.