Kimi K3
Moonshot AI's 2.8T-parameter open-weight multimodal reasoning model — 1M-token context, MXFP4, built for complex coding, knowledge work, and long-horizon agentic workflows.
Architecture
2.8T
MoE, Multimodal
Context Window
1M
Tokens
Intelligence
57
AA Index v4.0
License
Open
Open Weights
2.8T-Parameter Multimodal Reasoning
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 navigating large repositories, using tools, debugging, and iterating against images, logs, tests, and runtime feedback.
Its architecture uses KDA and Attention Residuals for computational efficiency, served in MXFP4 with a 1M-token context window. Released 27 Jul 2026.
Multimodal Agentic Flagship
Built for autonomous, multimodal agents
Long-Horizon Coding
End-to-end coding performance across Rust, Go, Python, front-end, DevOps, and performance optimization workflows.
Coding-Driven Design
Turns prompts and visual inputs into production-ready interfaces and lightweight full-stack workflows with structured layouts and visual polish.
Elevated Agent Swarm
Decomposes complex tasks into parallel, domain-specialized subtasks — scaling to large coordinated agent runs for end-to-end outputs.
Proactive Orchestration
Built for autonomous execution. Persistent background agents that manage schedules, execute code, and coordinate cross-platform operations with minimal oversight.
Frontier Coding & Agentic Performance
Artificial Analysis Intelligence Index v4.0 scores. Kimi K3 leads Moonshot's open-weight lineup on coding, tool use, and long-horizon agentic evaluations.
Intelligence Index
Better than 96% of models
GPQA Diamond
Better than 96% of models
τ²-Bench Telecom
Better than 95% of models
| Category | Benchmark | Score | Description |
|---|---|---|---|
| Reasoning | GPQA Diamond | 91% | Graduate-level scientific reasoning |
| Reasoning | Humanity's Last Exam | 36% | Humanity's Last Exam |
| Reasoning | τ²-Bench Telecom | 96% | AI agents in dual-control scenarios |
| Reasoning | AA-LCR | 70% | Long context reasoning evaluation |
| Reasoning | IFBench | 76% | Instruction-following accuracy |
| Reasoning | GDPval-AA | 49% | Agentic performance on real-world work tasks |
| Coding | SciCode | 53% | Python for scientific computing |
| Coding | Terminal-Bench Hard | 44% | Agentic coding & terminal use |
| Knowledge | AA-Omniscience Accuracy | 33% | Proportion of correctly answered questions |
| Knowledge | AA-Omniscience Non-Hallucination | 61% | Proportion of confidently answered questions that are correct |
Metrics sourced from Artificial Analysis and Moonshot's published evaluations. Reasoning (thinking) mode enabled.
Flexible Pricing Tiers
Choose the optimal balance of speed and cost for your workflow. Prices are per 1M tokens.
| Tier | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| Overnight (24H) | $1.50 | $7.50 |
| Async | $2.15 | $11.25 |
| Realtime | $3.00 | $15.00 |
Context window natively supported up to 1M tokens.
Start Building in Minutes
Kimi K3 is accessible via OpenAI-compatible endpoints.
from openai import OpenAI
client = OpenAI(
api_key="your-api-key-here",
base_url="https://api.doubleword.ai/v1"
)
# Long-horizon multimodal agentic task (thinking enabled by default)
response = client.chat.completions.create(
model="moonshotai/kimi-k3",
messages=[
{"role": "user", "content": "Plan and execute a 3-step refactor of this codebase."}
],
# To disable step-by-step reasoning:
# extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)
print(response.choices[0].message.content)💡 Pro Tip
K3 shines on long-horizon agentic work — let it sustain reasoning across planning, tool use, and iterative debugging against images, logs, tests, and runtime feedback. Route bulk repository or evaluation sweeps to the Async or Overnight tier to cut costs further.