HomeAI models
MoonshotAI: Kimi K3
Moonshot AI
•
Kimi K3 (Max)
•
Released 
July 2026

Moonshot AI
Kimi K3 (Max)

Open-weight 2.8T reasoning model with a 1M-token window and vision input for long-horizon coding agents.

Modality:
Text
Image
PDF
model ID
moonshotai/kimi-k3
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

38.91
tok/s

Intelligence Index *

43.6
/ 100

Context Window *

1048576
tokens

Input price

4.8
Anytoken

Output price

81
Anytoken
Kimi K3 (Max): Open-Weight Frontier Reasoning for Long-Horizon Coding Agents Kimi K3 is Moonshot AI's most capable flagship model, with 2.8 trillion parameters, native visual understanding and a 1M-token context window. It succeeds the K2 line. Reasoning is always on, and this Max configuration runs at the highest effort level. Moonshot targets long-horizon coding, knowledge work and reasoning. It fits repository-scale coding agents and multi-step research jobs where answer quality matters more than response speed. Access Kimi K3 (Max) via the AnyAPI.ai API

Performance

Frontier-Class Intelligence at Open-Weight Scale, Traded Against Speed

Kimi K3 (Max) scores 44 on the Artificial Analysis Intelligence Index, well above the median of 19 for comparable models. Output speed is about 40.6 tokens per second, an independent measurement. Combined with a 1M-token window, a large repository and its test logs can fit in one request. That makes this a quality-first model for hard, multi-step tasks rather than an interactive default. In production, budget for long reasoning traces and use it where a correct result outweighs response time.

Benchmarks

Independent Data: Strong Intelligence Score, Below-Median Speed

<cite index="23-3">Artificial Analysis rates Kimi K3 (Max) at 44 on its Intelligence Index, ahead of Claude Sonnet 5 (Max) at 38</cite>, and <cite index="19-3">GLM-5.2 (Max) at 34</cite>. <cite index="21-3">GPT-5.6 Sol (Max) scores 47.</cite> Output speed is roughly half that of Sol and Sonnet 5. <cite index="26-9">Speed varies by host: the fastest provider is 8.1x faster than the slowest, at 35.7 tokens per second.</cite> <cite index="24-10">Time to first answer token is about 50 seconds</cite>, because reasoning precedes the answer.

Output Speed

*
38.91
tok/s

Intelligence Index

*
43.6
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
94
%

HLE *

Adherence to multi-step structured instructions.
47
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

Kimi K3 (Max) accepts text and visual input and returns text. It has a 1,048,576-token context window shared between prompt and output. <cite index="29-2">max_completion_tokens defaults to 131,072 and can go up to 1,048,576</cite>, so set limits explicitly. Reasoning cannot be disabled; effort accepts low, high or max, with max as default. <cite index="36-7">Temperature, top_p and n are fixed</cite>, so tuning happens through reasoning effort and prompting rather than sampling.
Verified Specifications — 
Kimi K3 (Max)
*
Input modalities
Text
Image
PDF
output modalities
Text
Context window
1048576
 tokens
Maximum output tokens
1048576
Reasoning
Yes
Knowledge cutoff
July 2026
Pricing (standard)
4.8
 AnyTokens in
 / 
81
 AnyTokens out

Comparison

Kimi K3 (Max) or Kimi K2.7 Code for Coding Agents?

Both models come from Moonshot AI and target agentic coding, so K2.7 Code is the natural cheaper alternative for K3 users. <cite index="16-5">K3 has a 1,048,576-token window, against K2.7 Code's 262,144.</cite> <cite index="34-4">K3 supports reasoning_effort; K2.7 Code does not.</cite> <cite index="24-9">Artificial Analysis measures K2.7 Code at 52 tokens per second</cite>, faster than K3's low-40s. K2.7 Code also has much lower token prices. The decision is whether your task needs K3's extra reach.

Dimension
Kimi K3 (Max)
Kimi K2.7 Code
Context window *
1048576
tokens
262144
tokens
Output speed *
38.91
tok/s
74.19
tok/s
Intelligence Index *
43.6
25.8
Input pricing
4.8
AnyToken
4
AnyToken
Output pricing
81
AnyToken
20.4
AnyToken
Knowledge cutoff *
July 2026
June 2026

Choose Kimi K3 (Max) when repositories or evidence sets exceed 262K tokens, when agents run long autonomous sessions, or when screenshots and test output are part of the loop. Choose Kimi K2.7 Code for routine implementation, debugging and tool loops where lower token cost and faster output matter more than the larger window. <cite index="16-9,16-10">Test K3 on tasks beyond 256K context, and otherwise compare completed-task quality before changing defaults.</cite>

Limitations & Trade-offs

Where Kimi K3 (Max) falls short

1
Slow generation and late first answer. <cite index="20-17">Kimi K3 (Max) generates about 40.6 tokens per second, at the lower end of comparable open-weight models.</cite> <cite index="24-10">Time to first answer token is about 50 seconds</cite> in Artificial Analysis data. This rules it out for chat UIs, autocomplete and other interactive paths. Batch jobs, background agents and asynchronous pipelines tolerate it. For latency-sensitive features, a faster non-reasoning or lower-effort model is the better choice.
2
Reasoning overhead you cannot remove. <cite index="34-4">K3 always reasons and does not support the thinking parameter</cite>, and the default effort is max. <cite index="30-1">Thinking tokens are billed as output.</cite> <cite index="31-3">A tight token cap can spend the whole allowance on reasoning and return empty or truncated content.</cite> Short, high-volume calls such as classification or extraction waste budget here. Use low effort for them, or route them to a cheaper model.
3
Sensitivity to thinking history. <cite index="44-12,44-13">K3 was trained with preserved thinking, and generation quality can become unstable if a harness drops prior thinking or an ongoing session is switched over from another model.</cite> Agent frameworks that trim or strip reasoning_content will degrade results, and mid-session model swaps are risky. Test your orchestration layer before production. If your stack cannot preserve reasoning fields, a model without this dependency is safer.
4
Not the top of the intelligence ranking. <cite index="21-3">GPT-5.6 Sol (Max) scores 47 against K3's 44 on the Artificial Analysis Intelligence Index</cite>, and <cite index="45-2">Moonshot's own report concedes K3 still trails the most powerful proprietary models.</cite> Artificial Analysis also rates its token pricing as expensive and its token use as somewhat verbose. For the hardest tasks, or high-volume workloads where cost per task dominates, evaluate a stronger or cheaper alternative.

Best-Fit Workloads

Where this model earns its place

01

Repository-scale coding agents

‍
Moonshot positions K3 for long-horizon coding, and its 1,048,576-token window lets an agent hold large codebases, logs and test output in one session. Moonshot lists agent scenarios such as Kimi Code, Codex and Claude Code. It fits multi-file refactors and autonomous debugging. Preserve reasoning_content between turns, and expect slower responses than lighter coding models.

02

Vision-in-the-loop front-end development

‍
OpenRouter's model description says K3 iterates against images, logs, tests and runtime feedback.Wikipedia reports it outperformed competitors on Arena.ai's front-end web development benchmark at launch. Use it where an agent must inspect screenshots of a rendered UI and revise code. The per-request cost and speed profile make it better for build-time generation than interactive previews.

03

Long-horizon research and knowledge work

‍
Fireworks lists agentic deep research among K3's main uses, and Moonshot highlights end-to-end knowledge work such as research reports with interactive visualizations. With max-effort reasoning and long context, it suits multi-step investigations over large source sets. Latency is high, so run these as background jobs rather than live user requests.

04

Large-document and evidence analysis

‍
Moonshot lists flat per-token pricing with no tiering by context length, and context caching is automatic for regular requests. That helps when many questions run against the same large corpus, contract set or log archive. Send the stable prefix first so cache hits apply. Skip it for short documents, where reasoning overhead outweighs the benefit.

Pricing in anytokens via AnyAPI
Input
4.8
₳
Output
81
₳
Cache write
—
₳
Cache read
3.3
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsComparisonLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access Kimi K3 (Max) via AnyAPI.ai

Access Kimi K3 (Max) through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Kimi K3 (Max) without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access Kimi K3 (Max) and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test Kimi K3 (Max) against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use Kimi K3 (Max) from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use Kimi K3 (Max) for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

Kimi K3 offers a 1,048,576-token context window, which covers input and output together. max_completion_tokens defaults to 131,072 and can be raised to 1,048,576. Because the window is shared, long prompts reduce the room available for the answer and its reasoning trace.

No, not on Moonshot's official API. kimi-k3 always reasons and does not support the thinking parameter. You can set reasoning_effort to low, high or max, with max as the default. Some third-party hosts expose their own toggles, so check the behavior of the endpoint you use.

Artificial Analysis gives Kimi K3 (Max) an Intelligence Index score of 44, against a median of 19 for comparable models. GPT-5.6 Sol (Max) scores 47. Output speed is about 41 tokens per second, so K3 is strong on quality and slow on speed.

Moonshot describes native visual understanding, and developer guides report image and video input with text output only. Artificial Analysis lists text and image input, with text output. Video support can vary by endpoint, so test video requests on your integration before relying on them.

Moonshot states that K3 uses flat pay-as-you-go pricing with no tiering by context length. Cache writes, cache-hit input, cache-miss input and output are billed at separate per-token rates. Long prompts still cost more in total simply because they contain more tokens. Current pricing appears in the AnyTokens component.

* Benchmark data source: Artificial Analysis artificialanalysis.ai