k2 - Dark Side of the Moon Kimi's newest MoE Architecture Base Model

What's k2?

k2 is a MoE architecture base model with superb code and Agent capabilities from Moonshot AI, with 1T total parameters and 32B activation parameters. k2 model outperforms other mainstream open-source models in benchmark performance tests in the main categories of General Knowledge Reasoning, Programming, Mathematics, and Agent. k2 model context length is 128k, does not support visual features. It supports ToolCalls, JSON Mode, Partial Mode, and Networked Search.

Main functions of k2

• Superb code capability: Optimized for programming tasks, supporting complex code generation, debugging, interpretation and cross-language conversion.
• Agent capability: Supports multi-step ToolCalls to autonomously plan and execute task chains (e.g., data queries, API calls, file operations, etc.).
• Mathematics and logical reasoning: outperforms mainstream open-source models in mathematical competitions (e.g., AIME), logic puzzles, and scientific computation.

k2's official website address

• Official website address::Kimi Intelligent Assistant

How to use k2

• Visit kimi intelligent assistant: Visit the official website of Kimi Intelligent Assistant and choose to use the k2 model by default.
• Getting the API key: Register and login to the Moonshot AI Open Platform. Enter "API key" page, create and copy the key.

Technical characteristics of k2

• MoE Architecture: 1 trillion total parameters and 32 billion active parameters, balancing performance and efficiency.
• Context length: 128K tokens (about 250,000 Chinese characters), suitable for long document analysis or long conversations.
• nonvisual model: specializing in text processing.Does not support picture comprehension(need to be replaced by kimi-latest-vision).

Model pricing for k2

• cache hit: If the content of the request is already in the system cache, the input portion is pressed as ¥1.00/million tokens billing
• Cache misses: brand new or uncached content, the input portion presses the ¥4.00/million tokens billing
• output section: whether cached or not, uniformly press ¥16.00/million tokens billing
• Context length: Maximum support for a single request 131,072 tokens(≈250,000 characters)

Application Scenarios for k2

• Code and Software Development: K2 supports reading tens of thousands of lines of source code or the entire requirements document to generate a complete project skeleton.
• Intelligent Agents and Process Automation: K2 supports understanding of natural language commands and autonomous calls to databases, file systems, email or internal APIs to complete a multi-step business closure.
• Mathematical Reasoning and Research Assistance: Users can enter an entire paper, contest question, or complex formula at once, and the model will give step-by-step derivations, reproducible Python/JAX/PyTorch experiment scripts, and output LaTeX derivations that can be plugged directly into the paper.
• Text Insight: Legal, audit, and O&M teams can quickly complete protocol comparisons, compliance checks, or fault localization using the k2 model.