Loading page
FlagshipOpen SourceMoonshot
Kimi K3
Kimi K3 is Moonshot AI's flagship open-weight model in the Kimi family, combining a 1,000,000-token context window with text and image input and tool calling.
Context 1.0M
Tier Flagship
Tools Supported
License Open Source
Modalities text, image
Input from
$2.00 / 1M tokens
across 11 providers
API Pricing
Cheapest on EcoHash — 29% below avg| Provider | Input / 1M | Output / 1M | Cached / 1M |
|---|---|---|---|
| $2.00 | $12.00 | - | |
| $2.34 | $11.70 | $0.261 | |
| $2.55 | $12.95 | $0.285 | |
| $2.85 | $14.25 | $0.285 | |
| $3.00 | $15.00 | $0.750 | |
| $3.00 | $15.00 | - | |
| $3.00 | $15.00 | $0.300 | |
| $3.00 | $15.00 | - | |
| $3.00 | $15.00 | - | |
| $3.00 | $15.00 | $0.300 | |
| $3.00 | $15.00 | $0.300 | |
| $3.10 | $15.50 | $1.55 |
Prices updated daily. Last check: Sep 11, 2026
Kimi K3 pricing by provider
Input, output, and batch rates, plus alternatives, for one provider at a time.
Performance & Benchmarks
Source: Artificial Analysis →Intelligence
43.8 / 100
Coding
76.2 / 100
Output Speed
37.0 t/s
Latency (TTFT)
2.6s
Reasoning & Knowledge
- GPQA Diamond93.5%
- Humanity's Last Exam46.9%
Coding
- SciCode59.5%
Agentic & Tool Use
- Terminal-Bench v2.185.0%
- τ-bench Banking46.0%
Instruction & Long Context
- Long-Context Reasoning88.7%
Benchmarks measured Sep 2026. Scores are independent evaluations, not vendor-reported.
Model Details
General
- Creator
- Moonshot
- Family
- Kimi
- Tier
- Flagship
- Context Window
- 1.0M
- Modalities
- Text, Image
Capabilities
- Tool Calling
- Yes
- Open Source
- Yes
- Subtypes
- Chat Completion, Code Generation
- Aliases
- kimi-k3, moonshotai/kimi-k3, moonshotai-kimi-k3, kimi-k3-thinking
Strengths & Limitations
Strengths
- 1,000,000-token context window supports whole-repository and large document-set prompts without chunking
- Open weights allow self-hosting and serving by multiple independent inference providers
- Multiple hosting options typically produce a competitive spread of prices, visible in the pricing table on this page
- Accepts image input in addition to text, so visual documents and screenshots can be included in prompts
- Tool calling support enables integration with agent frameworks and function-dispatch pipelines
- Flagship tier of the Kimi family, with code generation listed among its target modality subtypes
- A kimi-k3-thinking alias indicates an available reasoning-oriented serving mode
Limitations
- Measured output throughput of roughly 40 tokens per second is modest for latency-sensitive interactive use
- Time to first token of about 2.1 seconds in independent testing adds noticeable startup delay to short requests
- Filling anything close to the 1M-token context is expensive in tokens and slow regardless of the per-token rate
- Behavior, quantization, and speed can differ between providers serving the same open weights
- We do not track published benchmark accuracy scores for this model, so quality comparisons rely on external evaluations
Key Features
•1,000,000-token context window
•Text and image input (multimodal prompts)
•Tool calling / function calling
•Open-weight release, self-hostable
•Chat completion interface
•Code generation capability
•Reasoning-oriented variant exposed as kimi-k3-thinking
•Available under multiple provider aliases including moonshotai/kimi-k3
About Kimi K3
Kimi K3 is a large language model from Moonshot AI (Moonshot) and sits at the flagship tier of the Kimi family. It is released with open weights, which means it can be served by multiple independent inference providers as well as accessed through hosted APIs — a distinction that usually produces a wider spread of prices and deployment options than closed models of comparable size. The model is exposed under several identifiers, including kimi-k3, moonshotai/kimi-k3, and kimi-k3-thinking, the last of which indicates a reasoning-oriented serving mode.
The headline technical characteristic is the 1,000,000-token context window, which places Kimi K3 among the small group of models that can accept very large document sets, entire repositories, or long conversation histories in a single request. Kimi K3 accepts both text and image input, and its modality subtypes are chat completion and code generation, indicating it is targeted at conversational assistants and programming work. Tool calling is supported, so the model can be wired into agent frameworks that dispatch function calls. Independent throughput measurements from Artificial Analysis put output generation at roughly 40 tokens per second with a time to first token of about 2.1 seconds; both figures vary by provider and by whether extended reasoning is enabled.
In practice, a model with this profile is used for long-context analysis, multi-file code work, and agentic pipelines where the ability to keep large amounts of source material in context reduces the need for retrieval plumbing. Because the weights are open, teams comparing Kimi K3 against closed flagship models typically weigh the option of self-hosting or choosing among multiple API vendors against the convenience of a single first-party endpoint. Current per-provider rates for Kimi K3 are shown in the pricing table on this page.
Common Use Cases
Kimi K3 suits workloads where context size is the binding constraint: reviewing or refactoring across many source files, analyzing long contracts, research corpora, or logs in a single pass, and maintaining long-running agent sessions that accumulate history. Tool calling makes it usable as the planning model in agentic systems, and image input lets it handle mixed documents such as specs with diagrams or UI screenshots. Its flagship position and code generation focus point toward development assistance and complex multi-step reasoning rather than high-volume classification or short-form extraction, where its ~40 tokens/second output rate and ~2.1 second time to first token make cheaper, faster small models a better fit. Because the weights are open, it is also a candidate for teams with data-residency or on-premise requirements that rule out closed API-only models.
Frequently Asked Questions
How much does Kimi K3 cost to use?
Pricing depends on which provider you use and the pricing type (input tokens, output tokens, cached input, and any long-context surcharges are often billed differently). Because Kimi K3 is an open-weight model served by several vendors, rates can vary considerably between them. See the pricing table on this page for current per-provider figures.
What is Kimi K3 best used for?
Long-context work and code-heavy tasks. The 1,000,000-token window suits whole-repository code analysis, large document review, and agent sessions with long histories, while tool calling supports function-dispatch workflows. It is less suited to high-volume, latency-sensitive short requests given its measured throughput.
Is Kimi K3 open source?
Our metadata records Kimi K3 as an open-weight release from Moonshot, which is why it appears under provider-prefixed identifiers such as moonshotai/kimi-k3 across multiple inference platforms. Check Moonshot's own license terms for the specific conditions on commercial use and redistribution.
What is the difference between kimi-k3 and kimi-k3-thinking?
Both are aliases for Kimi K3 in our database. The kimi-k3-thinking name denotes a reasoning-oriented serving mode that produces extended internal deliberation before answering, which generally increases output token counts and latency compared with the standard chat mode.
Can Kimi K3 process images?
Yes. Kimi K3's recorded modalities are text and image, so image input can be included alongside text in prompts. Whether a specific provider exposes image input on its endpoint should be confirmed against that provider's API documentation.
How fast is Kimi K3?
Independent measurements from Artificial Analysis report roughly 40.25 output tokens per second and a time to first token of about 2,139 ms. Actual figures depend on the provider, the hardware, the length of the prompt, and whether the reasoning mode is active.