MoonshotAI

Kimi K3.1 Pro

MoonshotToken-basedDynamic Pricing
Alias:kimi-k3.1-pro
Create API Key

Compare kimi-k3.1-pro API pricing, supported endpoints, capabilities and access options on Modelsell.

长上下文多模态推理编程工具调用结构化输出new
Starting price
View full pricing
Context
1M
Maximum input window
Max output
1M
Maximum tokens per response
Modalities
→

Pricing by Supplier

official
官方接口直连
Input/ 1M tokens
$2
Output/ 1M tokens
$10
Cache Read/ 1M tokens
$0.2
Cache Write (5m)/ 1M tokens
$2
Cache Write (1h)/ 1M tokens
$4

Capabilities / Supported modalities

StreamingFunction callingToolsJSON modeStructured outputReasoningVision
Input
Output

Provider & data privacy

Provider
Moonshot AIDocs
Tokenizer
Kimi tokenizer
License
Proprietary (commercial)Proprietary
Data retention66 daysNot used for upstream training by default

Performance

Benchmarks

Scores on standardized evaluations. Higher percentages are better — and rank percentile shows

Metrics sourced fromArtificial Analysis 2026-09-30·Kimi K3.1 Pro

This model is not included in the current benchmark snapshot.

Missing models or measurements are not zero scores.

Benchmark charts preserve the source model selection and reasoning settings. Missing models or measurements are not zero scores, and benchmark cost or speed is not this site’s service commitment.

About Kimi K3.1 Pro

把复杂材料推进到可检查的成果

Kimi K3.1 Pro 是月之暗面的 3T 级多模态推理模型,能够结合文字、图像和视频理解任务,并输出文字、代码或结构化数据。1,048,576 token 上下文适合同时阅读大型代码库、多份研究材料和较长的工作记录,减少反复切分资料的成本。

在编程任务中,可以先让它梳理模块和约束,再完成实现、测试与修正;处理 3D 项目时,结合场景图、空间关系和脚本要求开展建模与交互设计。金融研究则应提供数据口径、报告日期和原始材料,让模型区分事实、假设与推断,并用计算工具复核关键数字。

模型始终启用推理,可通过 reasoning_effort 选择 low、high 或 max,默认 max。较轻的任务可以先试 low,再依据结果调整;多步研究和复杂代码修改适合保留更充分的推理预算。图像和视频用于理解与分析,生成的图片、视频或 3D 文件需要由应用提供相应工具完成。

需要成本更低的长流程处理时可以选择 Pro;需要更快完成交互式任务时,可以选择同系列 Highspeed。长上下文并不替代验证:代码需要运行测试,图表需要对照数据,研究结论需要回到原始材料检查。

Use cases and prompting

大型代码项目示例

“阅读仓库说明、关键模块与报错截图,先定位调用链和根因,再修复问题。保留现有接口与数据格式,说明修改范围,运行相关测试,并记录仍未验证的路径。”

先提供项目目标、目录索引和关键约束,再补充相关源码与运行结果。通过工具调用执行测试后,完整保留返回的 assistant 消息(包括 reasoning_content)和工具结果,继续下一轮处理。

API 参数建议

使用 /v1/chat/completions,model 设为 kimi-k3.1-pro。reasoning_effort 支持 low、high、max,默认 max;不要传入旧版 K2 的 thinking 配置。temperature、top_p、n 和两项 penalty 使用固定值,建议省略。

max_completion_tokens 默认 131,072,参数上限 1,048,576;实际可用输出还受输入长度及整体上下文预算约束,推理 token 也会消耗输出预算。流式请求可启用 stream_options.include_usage 获取最终用量。

多轮对话及工具调用应原样保留完整 assistant 消息。需要严格结构化结果时,使用 response_format 的 json_schema,并只解析最终 message.content。图像建议使用 base64 数据,不直接依赖公共图片 URL。

固定资料与工具定义放在前缀,变化的提问和工具结果放在后部,更容易复用缓存。缓存命中、缓存写入与普通输入分别计费,输出用量包含推理部分。

API access

API documentation

Code samples

RequestPOST/v1/chat/completions
Example request
Parameters
ParameterTypeDefault / rangeDescription
temperature
number
=10 ~ 2
Sampling temperature; lower is more deterministic
top_p
number
=10 ~ 1
Nucleus sampling probability mass
max_tokens
integer>= 1Maximum number of tokens in the response
frequency_penalty
number
=0-2 ~ 2
Penalises repetition of frequent tokens
presence_penalty
number
=0-2 ~ 2
Encourages introducing new topics
stop
array—Up to 4 strings that stop generation
seed
integer—Deterministic sampling seed (best-effort)
n
integer
=1>= 1
Number of completions to generate
stream
boolean
=false
Stream tokens via Server-Sent Events
response_format
object—Force JSON object or schema-conforming output
tools
array—Tool / function declarations the model may call
tool_choice
string
autononerequired
Tool-choice policy or specific tool name
logprobs
boolean
=false
Return per-token log probabilities
top_logprobs
integer0 ~ 20Number of top log probabilities returned per token
logit_bias
object—Per-token logit bias map
user
string—End-user identifier for abuse monitoring

Replace <YOUR_API_KEY> with the API key from your token settings.

Authentication

All requests must include Authorization: Bearer <TOKEN> header. Anthropic-formatted endpoints accept the x-api-key header instead.

Generate tokens from the Tokens page; you can scope them to specific models, groups, IPs, and rate-limits.

Supported parameters

Generation parameters
ParameterTypeDefault / rangeDescription
temperature
number
=10 ~ 2
Sampling temperature; lower is more deterministic
top_p
number
=10 ~ 1
Nucleus sampling probability mass
max_tokens
integer>= 1Maximum number of tokens in the response
frequency_penalty
number
=0-2 ~ 2
Penalises repetition of frequent tokens
presence_penalty
number
=0-2 ~ 2
Encourages introducing new topics
stop
array—Up to 4 strings that stop generation
seed
integer—Deterministic sampling seed (best-effort)
n
integer
=1>= 1
Number of completions to generate
stream
boolean
=false
Stream tokens via Server-Sent Events
response_format
object—Force JSON object or schema-conforming output
tools
array—Tool / function declarations the model may call
tool_choice
string
autononerequired
Tool-choice policy or specific tool name
logprobs
boolean
=false
Return per-token log probabilities
top_logprobs
integer0 ~ 20Number of top log probabilities returned per token
logit_bias
object—Per-token logit bias map
user
string—End-user identifier for abuse monitoring

Rate limits

SupplierRPMTPMRPD
officialUnlimitedUnlimitedUnlimited

No restriction

Frequently asked questions about kimi-k3.1-pro

What is kimi-k3.1-pro?

Compare kimi-k3.1-pro API pricing, supported endpoints, capabilities and access options on Modelsell.

How do I call kimi-k3.1-pro?

Create an API key with access to kimi-k3.1-pro, then use the exact model ID and a supported endpoint from the API access section. Request fields depend on the selected endpoint.

How is kimi-k3.1-pro priced?

Pricing depends on the selected provider group and the model billing unit. The current input, output, request, or media prices are shown on this page before sign-up.

What is the context window of kimi-k3.1-pro?

The model catalog lists a context window of 1048576 tokens. Check the selected endpoint for request limits.

What is the maximum output of kimi-k3.1-pro?

The model catalog lists a maximum output of 1048576 tokens. Your request settings may set a lower limit.

How should I evaluate kimi-k3.1-pro for my project?

Start with the use cases and prompting guidance on this page, then evaluate the model with representative inputs from your project.