Moonshot AI
Kimi K3Open
Moonshot’s frontier MoE with 1M context, top-tier agentic benchmark results.
Intelligence
83.3 #48
AAQI 57.0
Output Speed
37 t/s
TTFT 3.06 s · ITL 27.0 ms/tok
Blended price (3:1)
$6 /1M
In $3/M · Out $15/M
Context
1.0M
≈ 2,500 A4 pages · max out 65,536
Quality Domain Radar
6-Axis Frontier Evaluation vs Category Average & Flagship
Kimi K3
Claude Mythos Preview (#1)
Average (68)
Radar axes are approximated from category scores where direct benchmark measurements are missing — see Methodology.
Strengths & Trade-offs
Where it wins
- Best agentic tool-use among open models
- Open weights at 1M context
- Record BrowseComp-style agentic search scores for its class
Trade-offs to watch
- Slower output (37 tok/s) than flash-class peers
- Trails Claude/GPT on SWE-bench Verified
Killer Use Cases
Agentic browsing & search
Long-horizon web research with tool use — the K-line's signature workload.
Telecom-style tool agents
τ²-Bench-class customer-service tool loops.
Open 1M-context RAG
Self-hosted long-document assistants.
Developer Quickstart & API
Set
MOONSHOT_API_KEY first — get an API keycurl https://api.moonshot.ai/v1/chat/completions \
-H "Authorization: Bearer $MOONSHOT_API_KEY" \
-H "content-type: application/json" \
-d '{
"model": "kimi-k3",
"messages": [{"role": "user", "content": "Refactor this function for O(n) time and explain the change."}]
}'import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.moonshot.ai/v1",
api_key=os.environ["MOONSHOT_API_KEY"],
)
resp = client.chat.completions.create(
model="kimi-k3",
messages=[{"role": "user", "content": "Refactor this function for O(n) time and explain the change."}],
)
print(resp.choices[0].message.content)const res = await fetch("https://api.moonshot.ai/v1/chat/completions", {
method: "POST",
headers: {
Authorization: `Bearer ${process.env.MOONSHOT_API_KEY}`,
"content-type": "application/json",
},
body: JSON.stringify({
model: "kimi-k3",
messages: [{ role: "user", content: "Refactor this function for O(n) time and explain the change." }],
}),
});
console.log((await res.json()).choices[0].message.content);Official Resources & Links
API deployment
Compare hosts & latency
Throughput, TTFT and cache pricing across Groq, Cerebras, Together, DeepInfra, Fireworks, Bedrock and Azure.
GroqCerebrasTogether AIDeepInfraFireworks AIAWS BedrockAzure AI
Scores by Source
Category Scores
Benchmark Results
Head-to-head Live Matchups
Live Specs & Architecture
Under the hood
Kimi K3 — undisclosed MoE
- Frontier open-weight MoE with 1M context
- Top-tier agentic tool-use — the K-line is built for long-horizon agents
- INT4-quantized checkpoints available for cheaper serving
Release date
Jul 15, 2026
Architecture
Kimi K3 — undisclosed MoE
Parameters
Not disclosed
License
Open Weights
Context Window
1,000,000 tokens
Max Output
65,536 tokens
Modalities
text + image
Cache Pricing
read $0.3 · write $3.75 /1M
Reasoning
Effort levels
Input Price
$3/M
Output Price
$15/M
Output Speed
37 t/s
Time to First Token
3.06 s
Data Updated
Aug 5, 2026
Hardware & Local Inference
Data: Convly AI (CC BY 4.0) ↗Min GPU
Multi-node cluster (e.g. 2× 8-GPU H200 141GB nodes)
Estimated monthly bill
Estimated monthly spend
$0.00
Calculating savings vs GPT-4o...
Cite & embed
GitHub Markdown badge
Loading badge markdown...BibTeX
@misc{llmpodium2026kimik3,
title = {Benchmark Analysis and Intelligence Evaluation of Kimi K3},
author = {LLMPodium Research},
year = {2026},
url = {https://llmpodium.com/models/kimi-k3}
}Embed Score Badge
Add a live Podium Score badge to your GitHub or Hugging Face README — it links back to the live ranking.
Preview
Markdown
Loading badge snippet...HTML
Loading badge snippet...More from Moonshot AI
Alternative Models
Closest alternatives to Kimi K3 based on capabilities, pricing, and benchmark scores.