Kimi K3 guide
Kimi K3 is Moonshot AI's flagship large language model, introduced in July 2026. It is a 2.8-trillion-parameter Mixture-of-Experts model with 104 billion active parameters and a 1,048,576-token (1M) context window. It accepts text, image, and video input and is available through Kimi.com, Kimi Work, Kimi Code, and the Kimi API (model ID kimi-k3). Its open weights were released on Hugging Face on July 27, 2026 under the custom Kimi K3 License. [S1, S2, S3, S19, S21]
Overview
Moonshot AI positions Kimi K3 as its flagship model for long-horizon coding, knowledge work, and deep reasoning. K3 always operates in thinking mode; reasoning effort can be set to low, high, or max (the default). Moonshot AI reports approximately 2.5x better scaling efficiency compared to Kimi K2. [S1]
The model weights are public on Hugging Face and the hosted model is served through the Kimi API with an OpenAI-compatible endpoint. This page collects the verified specifications and dated source checks for Kimi K3 in one place.
Key features
All capability descriptions are based on official Kimi and Moonshot AI sources, each with a dated access check.
Kimi K3 supports a context window of 1,048,576 tokens. The default output token limit is 131,072, configurable up to the full 1,048,576. Automatic context caching activates when prompts exceed 256 tokens, and Moonshot AI reports a cache hit rate above 90% in coding workloads. [S1, S2, S4]
K3 is a Mixture-of-Experts model with KDA and Attention Residuals (AttnRes): 2.8 trillion total parameters, 104 billion active parameters, and 896 experts with 16 activated per token. [S1, S19]
K3 natively supports visual understanding. It accepts text, image (PNG/JPEG via base64, suggested <= 4K resolution), and video (via file upload, suggested <= 1080p) as input, and outputs text. [S1, S2]
Moonshot AI reports K3 can sustain long-horizon software engineering sessions, navigating large codebases and orchestrating terminal tools with minimal human oversight. [S1]
K3 supports tool calling with tool_choice values of auto, none, and required, dynamic tool loading, structured output via response_format with json_schema, and Partial Mode, as documented in the K3 quickstart. [S2, S4]
The official moonshotai/Kimi-K3 Hugging Face repository is public and contains 118 files, including 96 Safetensors weight shards totalling approximately 1.56 TB, released under the custom Kimi K3 License. [S19, S21]
API reference
These fields come from the official Kimi API documentation and may change over time.
| Field | Value | Sources |
|---|---|---|
| API model ID | kimi-k3 | S2, S3 |
| Endpoint | OpenAI-compatible API at https://api.moonshot.cn/v1 | S2, S3 |
| Thinking / reasoning control | Thinking is always on. reasoning_effort can be set to low, high, or max; default is max. | S2, S4 |
| Context window | 1,048,576 tokens (1M) | S1, S2, S4 |
| Default max output tokens | 131,072; configurable up to 1,048,576 | S2 |
| tool_choice | Supports auto, none, and required | S2, S4 |
| Dynamic tool loading | Supported and documented in the K3 quickstart with a code example | S2 |
| Structured output | Supports response_format with json_schema and strict: true | S2 |
| Partial Mode | Supported and documented in the K3 quickstart with a code example | S2 |
| Context caching | Automatic. Moonshot AI reports a cache hit rate above 90% in coding workloads. | S1, S2 |
API pricing
Pricing below comes from official Kimi API documentation and Moonshot AI's K3 announcement blog, accessed July 24, 2026. Prices may change; check the current official pricing pages for the latest rates.
| Billing Category | CNY per 1M tokens | USD per 1M tokens |
|---|---|---|
| Input — cached | ¥2.00 | ≈ $0.30 |
| Input — uncached | ¥20.00 | ≈ $3.00 |
| Output | ¥100.00 | ≈ $15.00 |
CNY pricing is documented on the Kimi Platform pricing page. USD pricing was stated in Moonshot AI's K3 announcement blog. [S5, S1]
Open weights
Kimi K3 model weights were released on Hugging Face on July 27, 2026. The public moonshotai/Kimi-K3 repository contains 118 files — 96 Safetensors weight shards totalling approximately 1.56 TB, plus 4 vendor-reported evaluation metadata YAML files. The model card lists 2.8 trillion total parameters with 104 billion activated parameters, and a context length of 1,048,576 tokens. Official GPU, VRAM, and minimum hardware requirements were not specified in the release materials. [S19, S21]
The release uses the custom Kimi K3 License, which permits commercial use subject to conditions. Read the full license and the OpenK3 Open Weights guide for the pinned repository revision, download commands, framework notes, and license conditions before use. Read the Kimi K3 Open Weights guide.
K3 and K2.7 Code
Both kimi-k3 and kimi-k2.7-code, plus kimi-k2.7-code-highspeed, appear in the official Kimi API model list as of July 24, 2026. [S3]
The official sources reviewed for this page did not state that K3 replaces K2.7 Code. [S3]
For a side-by-side breakdown, see the Kimi K3 vs K2.7 Code comparison and the Kimi K2.7 Code guide.
Known limitations
| Limitation | Detail |
|---|---|
| Benchmarks are vendor-reported | All benchmark results come from Moonshot AI's own evaluation. A third-party composite (BenchLM BenchAlign v5.2, July 23, 2026) ranked K3 #4 of 200 with a score of 80.96 at Low confidence; it aggregates vendor-reported scores and does not independently test models. [S1] |
| Official local hardware requirements not specified | The release materials do not state minimum GPU, VRAM, or local hardware requirements. [S19] |
| Local open-weight video workflow unconfirmed | The model card mentions video and the hosted API documents video input, but a local open-weight video input workflow has not been verified. [S19, S2] |
| Local inference not performed by OpenK3 | The approximately 1.56 TB repository has not been downloaded and shard integrity has not been verified by this project. [S19] |
FAQ
Kimi K3 is Moonshot AI's flagship large language model, introduced in July 2026. It is a 2.8-trillion-parameter Mixture-of-Experts model with a 1,048,576-token context window. It accepts text, images, and video as input and is designed for long-horizon coding, knowledge work, and deep reasoning. [S1, S2]
Yes. Kimi K3 is accessible through Kimi.com, Kimi Work, Kimi Code, and the Kimi API with model ID kimi-k3. Its open weights are also publicly available from the official moonshotai/Kimi-K3 Hugging Face repository. [S1, S2, S3, S19]
The model ID currently listed in official documentation is kimi-k3, served through an OpenAI-compatible endpoint at https://api.moonshot.cn/v1. [S2, S3]
Kimi K3 supports a context window of 1,048,576 tokens (1M). The default output token limit is 131,072, configurable up to the full 1,048,576. Automatic context caching is built in. [S1, S2, S4]
Yes. The weights were released on Hugging Face on July 27, 2026 under the custom Kimi K3 License. Commercial use is permitted subject to conditions; read the full license before use. [S19, S21]
As of July 24, 2026, official API pricing is ¥2.00 cached input, ¥20.00 uncached input, and ¥100.00 output per 1M tokens in CNY, and approximately $0.30, $3.00, and $15.00 per 1M tokens in USD, as reported by Moonshot AI. Prices may change. [S5, S1]
The official sources reviewed did not state that Kimi K3 replaces Kimi K2.7 Code. As of July 24, 2026, both kimi-k3 and kimi-k2.7-code appeared in the official Kimi API model list. [S3]
Related
Official sources
Each source below includes its own access date.
https://www.kimi.com/blog/kimi-k3https://platform.kimi.com/docs/guide/kimi-k3-quickstarthttps://platform.kimi.com/docs/modelshttps://platform.kimi.com/docs/api/models-overviewhttps://platform.kimi.com/docs/pricing/chat-k3https://huggingface.co/moonshotai/Kimi-K3https://huggingface.co/moonshotai/Kimi-K3/blob/9f62e4e9fffbd0a83ddd60e1c209d828994b3569/LICENSEhttps://github.com/MoonshotAI/Kimi-K3