Kimi K3 by Moonshot AI: Features, Capabilities, and What It Can Do

Kimi K3 is Moonshot AI's latest frontier model — and the world's first open-source model in the 3-trillion-parameter class. Released in July 2026, it brings a 1M token context window, native multimodal input, and a reasoning mode that runs by default across every query. For developers and creators building AI-powered workflows, Kimi K3 represents a significant shift in what open-source models can handle.

This is a full breakdown of what Kimi K3 is, how it works, and where it fits relative to other frontier models available today.

Key Takeaways

• Kimi K3 has 2.8 trillion total parameters — the largest open-source model released to date
• Only 16 of 896 experts activate per query, keeping inference efficient despite the parameter count
• The 1M token context window supports full codebases, long documents, and video files in a single pass
• Thinking mode is always on — reasoning effort is configurable at low, high, or max
• The API is OpenAI SDK compatible, making it a drop-in replacement in existing pipelines
• Full model weights are scheduled for public release on July 27, 2026
• Pricing is flat pay-as-you-go with no context-length tiering and automatic caching for repeated queries

What Is Kimi K3?

Kimi K3 is a 2.8 trillion parameter open-source large language model developed by Moonshot AI, built for advanced reasoning, long-context processing, and agentic workflows across text, images, and video.

It's the first model in the open-source space to reach the 3-trillion-parameter class — a threshold previously occupied only by closed proprietary models. Moonshot AI released Kimi K3 via its developer platform in July 2026, with full model weights scheduled for public release on July 27, 2026.

Kimi K3 Key Features

2.8 Trillion Parameters

Kimi K3 holds 2.8 trillion total parameters — the largest parameter count of any open-source model released to date, and the first to cross the 3-trillion threshold. Parameter count alone doesn't determine capability, but at this scale combined with Kimi K3's MoE architecture, it means the model draws on a significantly broader knowledge representation than any previous open-source alternative.

1M Token Context Window

The context window determines how much information the model can hold and reference in a single pass. At 1 million tokens — approximately 750,000 words — Kimi K3 can process an entire large codebase, a book-length document, or a full session of conversation history without losing earlier context. For teams building content supply chains or knowledge workflows that involve large volumes of source material, this window size changes what's achievable in a single model call.

Native Multimodal Input

Kimi K3 processes text, local images, and video files natively — no separate vision model required. This matters for visual reasoning tasks like frontend UI interpretation, design file analysis, and frame-level video understanding. For workflows that involve image-to-video generation or visual content pipelines, Kimi K3 can hold visual context across an extended session rather than treating each image as an isolated input. Images are passed as base64 or via the ms:// protocol; video files go through the same API interface.

Sparse Mixture of Experts (MoE)

Rather than activating all 2.8 trillion parameters for every query, Kimi K3 routes each request through 16 of 896 specialized experts via the Stable LatentMoE framework. This sparse activation means the model delivers frontier-level capability at a fraction of the compute cost a dense architecture of the same size would require — faster inference and lower per-token cost without sacrificing depth on complex tasks.

Thinking Mode Always Enabled

Kimi K3's reasoning mode runs on every query by default — it doesn't need to be switched on separately. Reasoning effort is configurable at low, high, or max, controlling how deeply the model works through a problem before responding. How you prompt the model affects output quality at every reasoning level — low effort returns faster responses for straightforward queries, while max effort engages the full depth of the model's reasoning capacity for complex multi-step tasks. This is different from models that offer reasoning as a separate, more expensive tier — with Kimi K3, it's part of every call.

2.5x Scaling Efficiency Over Kimi K2

Kimi K3 delivers approximately 2.5x the overall scaling efficiency of its predecessor, Kimi K2. In practical terms, the same compute budget produces significantly more capability — a measure of how effectively the architecture converts training resources into model performance, not just a reflection of raw parameter count.

OpenAI SDK Compatible

Kimi K3's API is fully compatible with the OpenAI Python SDK. Switching an existing OpenAI-based pipeline to Kimi K3 requires two changes: the base URL (https://api.moonshot.ai/v1) and the model name ("kimi-k3"). For teams already running MCP servers and image prompting pipelines in Claude Code, Kimi K3 slots into the same toolchain without rebuilding the integration layer.

Structured JSON Output

Kimi K3 supports strict schema validation for structured output — useful for any workflow that needs the model to return data in a predictable, parseable format rather than free-form text. If you're already using JSON prompting for video generation, the same structured approach applies here: define the schema, pass it with the request, and the model validates its output against it before returning.

Partial Mode

Kimi K3 supports partial mode, allowing the model to continue generating from a text prefix you provide. This is useful for constrained generation tasks where output needs to start from a fixed point — relevant for templated content generation and structured data extraction workflows.

Automatic Context Caching

Repeated queries that share the same prefix benefit from automatic caching. Cache hits are billed at a lower input rate than cache misses, making long-context workflows more cost-effective over time. For any workflow that loads a fixed document or codebase at the start of each session, the savings compound across every subsequent query in that session.

How Kimi K3 Works

Kimi K3's architecture combines several components that address the two most common bottlenecks in large models: compute efficiency and long-context handling.

Sparse Mixture of Experts (MoE)

A model with 2.8 trillion parameters would be computationally prohibitive to run if all parameters activated for every query. Kimi K3 uses a Sparse Mixture of Experts architecture — specifically, the Stable LatentMoE framework — that routes each query through only 16 of 896 available experts. The result is that the model draws on a massive parameter pool while keeping per-query compute manageable. This is the same architectural principle behind DeepSeek V3 and Mixtral, but at a significantly larger scale.

Kimi Delta Attention (KDA)

Long-context models typically degrade in quality as sequence length increases — earlier information gets diluted as the context window fills. Kimi K3's Kimi Delta Attention is a hybrid linear attention mechanism designed to maintain smoother information flow across longer sequences. Combined with Attention Residuals (AttnRes), it supports the deeper architecture required for a model at this scale without the degradation that normally comes with longer contexts.

1M Token Context Window

1 million tokens is approximately 750,000 words — enough to hold a large codebase, a full legal document set, or several hours of video transcription in a single context window. Understanding how in-context learning works at this scale helps frame what Kimi K3 can actually do with that window. Kimi K3 pairs this with automatic context caching, which reduces cost on repeated queries where the prefix doesn't change. For workflows that query the same document or codebase repeatedly, the caching effectively makes long-context work more affordable over time.

What You Can Do With Kimi K3

Coding and Software Engineering

Kimi K3 is built for long-horizon coding tasks — work that requires holding the state of a large codebase across many steps without losing context. It integrates with terminal tools and external tool connections, supports large codebase comprehension, and can coordinate across multiple files and modules simultaneously. For teams working on complex engineering tasks that smaller models lose track of mid-task, the 1M context window and agentic design make a practical difference.

Visual Reasoning

Native multimodal input means Kimi K3 processes images and video files alongside text without a separate vision model. Moonshot AI specifically highlights visual reasoning for frontend engineering, game development, and CAD workflows — use cases where the model needs to interpret a screenshot, wireframe, or design file as part of a task rather than just reading a text description of it.

Long Document Processing

The 1M token context window makes Kimi K3 practical for knowledge work that involves large volumes of source material — research synthesis, legal document review, technical documentation, or any task where the relevant information is spread across a long source. Automatic caching means documents loaded into context don't incur full input costs on every follow-up query.

Agentic Workflows

Kimi K3's thinking mode runs by default — it reasons through problems before responding, and that reasoning is configurable. At max effort, it applies extended reasoning to complex multi-step tasks. At low effort, it responds faster for queries that don't need deep analysis. This configurable reasoning makes it well-suited for production-oriented agentic AI workflows where some tasks need full reasoning depth and others don't. For a broader look at what AI agents can do across different platforms, the use cases extend well beyond coding into content production and creative workflows. If you want to run AI agents without building a pipeline from scratch, ImagineArt Computer lets you work with AI agents directly in the browser — no setup required. You can access the agent dashboard to start a session and test how a model like Kimi K3 handles multi-step tasks.

Kimi K3 vs. Other Frontier Models

Kimi K3's main differentiators in this comparison: the largest disclosed parameter count by a significant margin, open-source weights, and thinking mode enabled by default rather than as an optional switch. Gemini 2.5 Pro matches the 1M context window but is closed-source and carries context-length-based pricing. Llama 3.1 405B is the closest open-source alternative but at roughly one-seventh the parameter count and without native multimodal support.

How to Get Started With Kimi K3

1. Go to platform.kimi.ai/console/api-keys and generate an API key
2. Install the OpenAI Python SDK — requires Python 3.9 or above
3. Set the base URL to https://api.moonshot.ai/v1 and the model to "kimi-k3"
4. Set reasoning effort via the API parameter — "low", "high", or "max" depending on the task. If you're new to structuring prompts for reasoning models, this prompt engineering guide covers the core principles
5. Test your first query in the playground at platform.kimi.ai/playground before integrating into your pipeline

The OpenAI SDK compatibility means any existing code using the OpenAI client requires only two changes: the base URL and the model name. Streaming, structured JSON output, tool calling, and vision input all work within the same SDK interface.

Kimi K3 Pricing

Kimi K3 uses flat pay-as-you-go pricing with no tiering based on context length — a meaningful difference from models that charge more per token as the context window fills. Input tokens (both cache hits and cache misses) and output tokens are billed at uniform per-token rates.

Automatic context caching reduces effective input costs on workflows that repeatedly query the same prefix — relevant for any use case that loads a fixed document or codebase at the start of each session.

Exact token pricing is available at platform.kimi.ai.

Try ImagineArt

Access the model at platform.kimi.ai — the playground lets you test queries before building into a pipeline, and the API is available immediately with a free API key. If you're building workflows that include AI image generation alongside your Kimi K3 setup, ImagineArt's API connects directly to Claude Code and Codex CLI for image generation inside the same pipeline.

Frequently Asked Questions

What is Kimi K3?

Kimi K3 is a 2.8 trillion parameter open-source large language model developed by Moonshot AI. It supports text, image, and video input, includes a 1M token context window, and runs with thinking mode enabled by default. It's the first open-source model to reach the 3-trillion-parameter class.

Who made Kimi K3?

Kimi K3 was developed by Moonshot AI, a Chinese AI research company. The model was released via their developer platform in July 2026, with full open-source weights scheduled for public release on July 27, 2026.

Is Kimi K3 open source?

Yes. Kimi K3's model weights are being released publicly. As of July 20, 2026, the model is accessible via Moonshot AI's API platform, with full weight release scheduled for July 27, 2026.

What is the context window of Kimi K3?

Kimi K3 supports a 1M token context window — approximately 750,000 words. It includes automatic context caching for repeated queries with unchanged prefixes, which reduces input costs on workflows that reuse the same source material.

How does Kimi K3 compare to GPT-4o?

Kimi K3 has a significantly larger parameter count (2.8T vs. undisclosed for GPT-4o), a larger context window (1M vs. 128K tokens), and is open source. GPT-4o is a closed model with a shorter context window. Kimi K3's thinking mode is always enabled; GPT-4o's reasoning capabilities are available through separate models (o3/o4) rather than built into the base model.

Can Kimi K3 process images and video?

Yes. Kimi K3 has native multimodal capabilities — it processes local images and video files alongside text input. Images and video can be passed as base64 or via the ms:// protocol through the API.

How do I access Kimi K3 via API?

Kimi K3 is accessible through Moonshot AI's API at https://api.moonshot.ai/v1. It's compatible with the OpenAI Python SDK — set the base URL and model name to "kimi-k3" and existing OpenAI-based code works without further changes.

What is Kimi K3 best used for?

Kimi K3 performs best on long-horizon coding tasks, agentic multi-step workflows, visual reasoning with images and video, and long document processing. The 1M context window makes it practical for tasks where the source material exceeds what other models can hold in a single pass.