🚀 Kimi K3 is live - 2.8T params · 1M context · #1 Frontend Code Arena · Open weights July 27 · Try the API now →
✦ Try Kimi K3 Free →
RELEASED · JULY 16, 2026 · BY MOONSHOT AI

Kimi K3
Is Here

The world's largest open-weight AI model - 2.8 trillion parameters, 1 million token context, native vision and video, and #1 on the Frontend Code Arena. Available now via API and kimi.com. Open weights shipping July 27, 2026.

2.8T Parameters 1M Token Context #1 Frontend Code Arena KDA Hybrid Attention · 6.3× faster at 1M ctx Open Weights July 27 · Modified MIT
🔓 Open Weights Release
July 27, 2026
Full 2.8T weights · Modified MIT License · github.com/MoonshotAI/Kimi-K3
First "open 3T-class model" · K3 technical report ships same day
2.8TTotal Parameters
896Total Experts (16 active)
1MContext Tokens
#1Frontend Code Arena
6.3×Faster at 1M ctx (KDA)
VentureBeat: "Largest open-source model ever, rivaling top U.S. systems" Fortune: "Pushes Chinese AI into Fable-level territory" Bloomberg: "Closes gap with US rivals" Axios: "Fueling awe across the AI world - and alarm in Silicon Valley" #1 LMArena Frontend Code Arena - 1,679 Elo · 17-place jump from K2.6 93.5% GPQA Diamond - best open-weight result ever BrowseComp 91.2% · SWE Marathon 42.0 · Terminal-Bench 88.3% Cost per task: ~$0.94 - half the price of Opus 4.8 ($1.80) VentureBeat: "Largest open-source model ever, rivaling top U.S. systems" Fortune: "Pushes Chinese AI into Fable-level territory" Bloomberg: "Closes gap with US rivals" Axios: "Fueling awe across the AI world - and alarm in Silicon Valley" #1 LMArena Frontend Code Arena - 1,679 Elo · 17-place jump from K2.6 93.5% GPQA Diamond - best open-weight result ever BrowseComp 91.2% · SWE Marathon 42.0 · Terminal-Bench 88.3% Cost per task: ~$0.94 - half the price of Opus 4.8 ($1.80)
// 01 · ARCHITECTURE

Stable LatentMoE +
Kimi Delta Attention

K3 is built on two architectural innovations that separate it from every prior Kimi model: a new sparse MoE layout called Stable LatentMoE with 896 experts, and Kimi Delta Attention (KDA) - a hybrid linear-attention mechanism that delivers up to 6.3× faster decoding at 1M token context lengths.

// KIMI DELTA ATTENTION - TOKEN FLOW
Input Token
Route to 16 of 896 experts
KDA Layer
Linear attention + delta residual
Full Attn
Sparse global attention passes
Output
Recombined via shared experts
2.8TTotal Params
~50BActive per Token
896Total Experts
16Experts Active
1MContext Tokens
6.3×Faster Decode @ 1M
  • Stable LatentMoE - K3's new MoE layout with 896 total experts. Only 16 routed experts activate per token (~50B active parameters). Shared experts run in parallel to every token, providing consistent base capability while routed experts specialize. The "Stable" prefix refers to training stability - K3 is the first Moonshot model trained without using MuonClip at this scale, indicating architectural maturity.
  • Kimi Delta Attention (KDA) - A hybrid linear-attention mechanism with attention residuals. Standard transformers scale quadratically with context; KDA uses a linear recurrence for most context, inserting sparse full-attention "delta" passes at key intervals. Result: 6.3× faster decoding at 1M tokens compared to standard attention. A 1M token context is no longer theoretical - it's actually usable at these speeds.
  • Native multimodal - K3 natively accepts text, images, and video (frame-by-frame). No separate encoder step visible to the user. Ask about a chart, screenshot, diagram, or video frame alongside code or documents. Video understanding enables temporal reasoning ("what changes between frame 3 and frame 7?").
  • Always-on thinking - effort dial - Unlike K2.6 which had switchable thinking mode, K3 always reasons before responding. The interface replaces on/off with a reasoning_effort dial. At launch, the dial ships locked to maximum. Lighter and heavier modes are promised in a post-launch update.
  • Two configurations at launch - K3 Max for general chat, coding, and knowledge work. K3 Swarm Max for large-scale parallel agent processing - multi-agent workflows at K3-quality output.
  • Mooncake serving - 90%+ cache rates - Moonshot's Mooncake inference infrastructure maintains coding cache hit rates above 90%, which means the effective cached-input price ($0.30/1M) applies to most real-world coding workloads rather than the $3.00/1M uncached rate. Real cost per coding task is substantially lower than the headline rate suggests.
// 02 · BENCHMARKS

Performance at the Frontier

K3 doesn't win every benchmark - and Moonshot says so plainly in its launch post. It trails Claude Fable 5 and GPT-5.6 Sol overall. What it does deliver is genuine SOTA performance in specific high-value lanes: coding, frontend, long-horizon agentic work, and document-scale knowledge tasks.

🏆 #1 Overall
1,679 Elo
LMArena Frontend Code Arena
17-place jump from K2.6 (#18)
6/7 frontend domains · Only model to beat Fable 5 here
🧬 GPQA Diamond
93.5%
Best open-weight result ever published on GPQA
PhD-level science reasoning benchmark
🤖 SWE Marathon
42.0
K3's most significant agentic claim
vs Opus 4.8: 40.0 · Fable 5: 35.0
Long-horizon agentic software engineering
Important caveat on benchmarks Most benchmarks above come from Moonshot's own launch materials at "max" reasoning effort. Independent Artificial Analysis testing places K3 4th overall on their Intelligence Index (score: 57), behind Claude Fable 5 (~60) and GPT-5.6 Sol (~59). The hallucination rate rose to ~51% (up from K2.6's ~33%). Raw accuracy improved from 33% to ~46%. For factual and research tasks, keep verification in the loop.
// 03 · COMPARISON

K3 vs Fable 5 vs GPT-5.6 Sol vs Opus 4.8

K3 beats Opus 4.8 and GPT-5.5 outright. It wins specific benchmarks against Fable 5 and GPT-5.6 Sol while trailing them on overall intelligence. Open weights, price, and the scale gap make it a different kind of story than a benchmark table can fully capture.

Kimi K3Moonshot AI · Jul 16, 2026 Claude Fable 5Anthropic · 2026 GPT-5.6 SolOpenAI · 2026 Claude Opus 4.8Anthropic · 2026 GPT-5.5OpenAI · 2026
Architecture
Total parameters2.8T MoE~unknown~unknown~200B~200B
Open weights✓ Jul 27 · Mod. MIT
Context window1M tokens200K1M200K1M
Native multimodalText + Image + VideoText + ImageText + ImageText + ImageText + Image
Always-on thinking✓ effort dialPartial
Benchmark performance
AA Intelligence Index57 #4~60 #1~59 #2~56~51
LMArena Frontend Code1,679 Elo #11,612 Elo1,589 Elo1,531 Elo~1,480
GPQA Diamond93.5% Best open~96%~95%~89%~84%
SWE Marathon42.0 SOTA35.037.240.028.0
Terminal-Bench 2.188.3 #286.188.8 #182.474.2
BrowseComp91.2% SOTA88.4%87.6%83.2%74.8%
Program Bench77.8 SOTA77.177.674.368.9
AA Coding Index76.24 #174.873.269.161.4
AA-Briefcase Elo1,547 #2~1,620+~1,590~1,480~1,390
Cost & speed
API input / 1M tokens$3.00 (or $0.30 cached)$50.00+~$15.00~$15.00~$5.00
API output / 1M tokens$15.00~$50.00~$60.00~$75.00~$30.00
Cost per task (typical)~$0.94~$3.20~$1.04~$1.80~$1.20
Decoding at 1M context6.3× KDA boostStandardStandardStandardStandard

Competitor figures at max reasoning effort where comparable. Independent AA scores. Fable 5 / GPT-5.6 Sol / Opus 4.8 scores from Artificial Analysis and Moonshot's published launch comparison. All figures current as of July 17, 2026.

// 04 · API & PRICING

API Access & Pricing

Kimi K3 is live now on the Kimi API. OpenAI and Anthropic SDK compatible - change two lines from your existing Claude or GPT setup. Mooncake serving infrastructure keeps cache hit rates above 90% on coding tasks, making the effective cost far lower than the headline rate.

Cached Input 90%+ hit rate on code
Cache hits via Mooncake serving - most coding requests
$0.30
per 1M tokens
Fresh Input (Cache Miss)
First-time prompts, uncached document context
$3.00
per 1M tokens
Output Tokens
All generated tokens including thinking trace
$15.00
per 1M tokens
Web Search
Per call when K3 invokes web search tool
$0.015
per call
Real cost comparison Cost per typical task: K3 ~$0.94 vs GPT-5.6 Sol ~$1.04 vs Claude Opus 4.8 ~$1.80 vs Claude Fable 5 ~$3.20+. Cache rates above 90% on coding workflows mean the effective K3 rate is much closer to $0.30/1M input than $3.00/1M in practice.

Context access by plan

Model ID: kimi-k3 (Kimi Code: /model k3)
Moderato ($19/mo): 256K context max
Allegretto ($39/mo): Full 1M context
Direct API: Full 1M with max_tokens: 1048576

Python · OpenAI SDK · Kimi K3
# pip install openai from openai import OpenAI client = OpenAI( api_key="YOUR_KIMI_API_KEY", base_url="https://api.moonshot.ai/v1" ) # ── Kimi K3 - always-on thinking ── response = client.chat.completions.create( model="kimi-k3", max_tokens=32768, # reasoning_effort: coming in post-launch update messages=[ {"role": "system", "content": "You are an expert software architect."}, {"role": "user", "content": "Design a distributed rate-limiter for 100K req/s"} ] ) # Access K3's thinking trace print(response.choices[0].message.reasoning_content) print(response.choices[0].message.content) # ── K3 with full 1M context (Allegretto+) ── response = client.chat.completions.create( model="kimi-k3", max_tokens=1048576, # full 1M window messages=[ {"role": "user", "content": "Analyze this entire 800K-token codebase..."} ] ) # ── K3 Swarm Max for agentic pipelines ── response = client.chat.completions.create( model="kimi-k3-swarm-max", max_tokens=32768, messages=[{"role": "user", "content": "Orchestrate 20 agents to audit this monorepo"}] ) # ── Kimi Code CLI - switch to K3 ── # In terminal: /model k3 (or --model k3 at startup)
// 05 · 1M CONTEXT

What 1 Million Tokens Actually Means

📦

Entire Codebases

1M tokens holds ~750,000 lines of code. Load an entire monorepo, all dependencies and configuration files, and have K3 reason about the full system at once - no chunking, no retrieval overhead.

📚

Book-length Documents

~750,000 words per session - roughly 10 full novels or 3,000 research papers. Legal due diligence, multi-document contract analysis, regulatory filing review in one pass.

🎬

Video Frame Analysis

Frame-by-frame video reasoning within the 1M token window. Ask temporal questions about what changes across hundreds of frames. Unique to K3 in the Kimi lineup.

🤖

Long-Horizon Agents

Extended multi-turn agentic sessions that maintain full context across hundreds of tool calls without forgetting early decisions. K3's 1M window is why SWE Marathon scores beat models with 200K context.

6.3× Faster at 1M

KDA hybrid attention makes the 1M window actually usable - not just a number on a spec sheet. Standard quadratic attention at 1M context is impractically slow; KDA's linear recurrence changes that.

🏢

Enterprise Contexts

Long-form financial models, annual reports, multi-year regulatory submissions. K3 can hold an entire year's worth of enterprise documentation in a single context and reason across it all.

// 06 · VIDEOS & NEWS COVERAGE

Watch Kimi K3 in Action

Official Kimi K3 Demo - Frontend Code Arena #1 Performance

Kimi K3 Benchmark Comparison - Full Scorecard

K3 + Kimi Code - Full Codebase Analysis Session

K3 Open Weights - Self-Hosting Guide (July 27)

// 07 · HOW TO USE

Four Ways to Access Kimi K3

01

kimi.com / App

K3 is live at kimi.com and the Kimi mobile app. Select K3 from the model switcher. Full 1M context requires Allegretto plan ($39/mo). Moderato users get 256K. Free users get basic access. Native image and video upload supported in all tiers.

02

Kimi API

Model ID: kimi-k3 and kimi-k3-swarm-max. Base URL: https://api.moonshot.ai/v1. OpenAI and Anthropic SDK compatible. Get your key at platform.kimi.ai. Pricing: $0.30 cached / $3.00 uncached input · $15.00 output.

03

Kimi Code CLI

Switch to K3 in an active Kimi Code session: type /model k3. Or start with kimi --model k3. K3 is Moonshot's most capable coding model but uses more tokens per session. Best for complex long-session work where quality matters more than cost.

04

Self-Host (July 27)

Full 2.8T weights release on July 27 under Modified MIT license. Repo: github.com/MoonshotAI/Kimi-K3. Deployment requires A100-class GPU clusters. Supported engines: vLLM, SGLang, TensorRT-LLM. Technical report ships on the same date. Commercial use permitted.

// 08 · MODEL TIMELINE

The K-Series Journey: K2 → K3

K3
JULY 16, 2026 - LIVE NOW
Kimi K3 - 2.8T · 1M Context · #1 Frontend Code
Stable LatentMoE · 896 experts · KDA hybrid attention · Native video · Always-on thinking · Open weights July 27 · $3/$15 per 1M
World's largest open-weight model
JUNE 15, 2026
Kimi K2.7 Code HighSpeed - 260 tok/s · 6× faster
Same K2.7 model, extreme throughput serving. Up to 260 tok/s peak. Rolling out to Beta users.
JUNE 12, 2026
Kimi K2.7 Code - MCP Mark #1 · −30% reasoning tokens
81.1 MCP Mark Verified · Mandatory thinking · 256K ctx · Kimi Code default model.
APRIL 20, 2026
Kimi K2.6 - 300-Agent Swarm · 262K ctx · SWE-bench 80.2%
Claw Groups · Document-to-Skill · BrowseComp 86.3% · Kimi Work desktop agent launched on K2.6.
JANUARY 27, 2026
Kimi K2.5 - Native Multimodal · 100 Agents · AIME 96.1%
MoonViT 400M encoder · 256K ctx · 100 sub-agents / 1,500 steps · 4.5× speedup
NOVEMBER 2025
Kimi K2 Thinking - 300 tool calls · INT4 QAT · 2× speed
Interleaved reasoning + tool use. Tencent CodeBuddy integration. Native INT4 quantization via QAT.
JULY 2025
Kimi K2 - Flagship Open-Source · 1T MoE · 128K
The original 1T parameter release. SWE-bench 65.8% · MMLU-Pro 73.3% · τ²-bench 80% · Modified MIT.
// 09 · FAQ

Kimi K3 - Frequently Asked Questions

The World's Largest
Open-Weight AI - Live Now

Kimi K3: 2.8T parameters, 1M context, #1 Frontend Code Arena, open weights July 27. Try the API today or access via kimi.com. Free to start.