Agent and tool-use research
Studying an influential open-weight model designed around function calling and multi-step application workflows.
Independent tool overview
Kimi K2 at a glance
Kimi K2 is Moonshot AI's 2025 mixture-of-experts language model for coding, tool use, knowledge, and agentic workflows. Its original Base and Instruct weights remain downloadable, but Moonshot's current API documentation now recommends newer Kimi models for production use.
Visit the official Kimi K2 site ↗
- Developer
- Moonshot AI
- Architecture
- 1T-parameter MoE; 32B active
- Original context
- 128K tokens
- License
- Modified MIT License
Overview
What Kimi K2 is
Kimi K2 has one trillion total parameters while activating 32 billion for each token. The original release includes a Base checkpoint for research and customization and an Instruct checkpoint for chat, coding, and tool-using applications.
The original Instruct model is a fast, non-thinking model with a 128K context configuration. Moonshot later published an updated 0905 checkpoint with stronger agentic coding and 256K context, followed by separate K2 Thinking and newer multimodal Kimi generations.
The K2 weights remain available under a Modified MIT License, but hosting a model of this size is a serious infrastructure project. Moonshot's current API quickstart recommends Kimi K3, K2.7 Code, or K2.6, so new deployments should benchmark those successors before standardizing on the original K2.
Use cases
Who Kimi K2 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Coding evaluations
Testing code generation and repository tasks where generated changes can be executed and scored automatically.
Model customization
Research teams with substantial infrastructure that want a Base checkpoint for fine-tuning or specialized post-training.
Legacy K2 deployments
Organizations already using the original checkpoint that need accurate documentation before deciding whether to migrate.
Capabilities
Core Kimi K2 features
Mixture-of-experts design
Uses one trillion total parameters but activates 32 billion per token to balance model capacity with inference efficiency.
Base and Instruct checkpoints
Provides a foundation checkpoint for customization and a post-trained checkpoint for general chat and agentic applications.
Agentic optimization
Moonshot trained and evaluated K2 for tool use, coding, and autonomous problem-solving rather than only conversational response quality.
128K original context
The original Kimi-K2-Instruct configuration supports a 131,072-token context window.
Updated 0905 checkpoint
A later K2 Instruct variant expands context to 256K and targets stronger agentic coding behavior.
OpenAI-compatible serving
Official examples show local chat through an OpenAI-style client once a compatible inference server is running.
Multiple inference engines
Moonshot documents deployment paths involving vLLM, SGLang, KTransformers, and TensorRT-LLM.
Block-FP8 weights
The official checkpoints are distributed in block-FP8 form to reduce the footprint compared with full-precision weights.
Process
How the Kimi K2 workflow works
Step 1
Choose the right generation
Compare original K2 with K2.6, K2.7 Code, K3, and any hosted options before committing to a legacy checkpoint.
Step 2
Review the modified license
Check the Modified MIT terms, including the attribution requirement for very large commercial products or services.
Step 3
Select a checkpoint
Use Base only when custom training is justified; use Instruct or a newer successor for application-facing tasks.
Step 4
Plan the infrastructure
Size storage, accelerator memory, quantization, context, throughput, redundancy, and serving software for the real workload.
Step 5
Evaluate tools and code
Run representative tasks with executable tests, tool-call validation, latency measurement, and adversarial failure cases.
Step 6
Add production controls
Protect credentials and tools, restrict permissions, validate arguments and outputs, log actions, and require human approval for consequential changes.
Cost
Kimi K2 pricing and free plan
Moonshot publishes the original Kimi K2 checkpoints for download under a Modified MIT License. The weights do not carry a subscription fee, but users pay for compute, storage, engineering, security, and operations. Current hosted API pricing should be checked against Moonshot's live model catalog because the original K2 is no longer its recommended starting model.
Open-weight checkpoints
Free download
Download Kimi K2 Base or Instruct and operate it under the Modified MIT License.
- Self-hosted infrastructure required
- Base and Instruct variants
- Block-FP8 checkpoints
- Modified license terms apply
Pricing checked . Check current pricing at the source ↗
Assessment
Kimi K2 strengths and limitations
Where it stands out
- High-capacity mixture-of-experts model with a smaller active parameter count
- Weights, model card, code, deployment guidance, and technical report are public
- Designed explicitly for coding and tool-using workflows
- Several inference-engine paths are documented
- An updated K2 checkpoint extends context and agentic coding capability
What to consider
- The original K2 is no longer Moonshot's recommended default; current documentation points developers to newer Kimi generations.
- A one-trillion-parameter checkpoint remains expensive and operationally demanding despite sparse activation and FP8 weights.
- The original Instruct variant is a non-thinking model and differs materially from K2 Thinking and later multimodal models.
- Tool-using agents can take unsafe or incorrect actions unless permissions, arguments, outputs, and approval gates are tightly controlled.
- Moonshot's benchmark results should be reproduced on the buyer's own tasks, runtime, quantization, and context settings.
Compare
Kimi K2 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Consumer
Kimi K2.6
Kimi K2.6 is Moonshot's newer general-purpose model with multimodal input, thinking and non-thinking modes, and current API documentation.
Explore Kimi K2.6 →Coding
Kimi K2.7 Code
Kimi K2.7 Code is the newer Kimi option for coding agents and repository work.
Explore Kimi K2.7 Code →Consulting
Kimi K3
Kimi K3 is Moonshot's current flagship and the default recommendation in its latest API quickstart.
Explore Kimi K3 →Questions
Kimi K2 FAQs
What is Kimi K2?
Kimi K2 is Moonshot AI's 2025 mixture-of-experts language model for knowledge, coding, tool use, and agentic tasks. It has one trillion total parameters and activates 32 billion per token.
Is Kimi K2 open source?
Moonshot publishes the K2 code and model weights under a Modified MIT License. The license is permissive but includes an extra display requirement for products or services above specified scale thresholds.
Is Kimi K2 still current?
The original weights remain available, but they are no longer Moonshot's recommended starting point. Its current API guide highlights Kimi K3, K2.7 Code, and K2.6.
How much does Kimi K2 cost?
The downloadable weights have no subscription price. Self-hosting requires significant compute and operational spending, while hosted model prices should be checked in Moonshot's current platform catalog.
What is the difference between Kimi K2 and Kimi K2 Thinking?
The original K2 Instruct checkpoint is a reflex-style non-thinking model. Kimi K2 Thinking is a separate later checkpoint designed for explicit multi-step reasoning and longer tool-use sequences.
Bottom line
Our Kimi K2 verdict
Kimi K2 is still useful as an open-weight reference model and for teams maintaining an existing deployment. It should not be the automatic choice for a new build in 2026: Moonshot has released multiple successors and now recommends them in its own API guide. Benchmark the current Kimi lineup first, then use original K2 only when its specific compatibility, checkpoint, or research value wins.
Visit Kimi K2 website ↗Similar Agents Tools
View all Agents AI tools →
Comet
Comet - Perplexity’s new AI-first browser
ChatGPT Agent
ChatGPT Agent - ChatGPT now thinks and acts, proactively choosing from a toolbox of agentic skills to complete tasks for you using its own computer.
Mistral AI
Mistral AI: Develops open-weight, high-performance language models designed for efficiency and customization.
Action Agent
Action Agent - Writer's enterprise-ready autonomous AI that works on your behalf
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.