Kimi K2: Open Agentic Intelligence

Overview

Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model with 32 billion activated parameters and 1 trillion total parameters. It excels in frontier knowledge, reasoning, and coding tasks, optimized for agentic capabilities.

Key Features

• Large-Scale Training: Trained on 15.5T tokens with zero instability.
• MuonClip Optimizer: Novel optimization for stable scaling.
• Agentic Intelligence: Designed for tool use, reasoning, and autonomous problem-solving.

Model Variants

• Kimi-K2-Base: Foundation model for fine-tuning.
• Kimi-K2-Instruct: Post-trained for general-purpose chat and agentic tasks.

Technical Specifications

• Architecture: Mixture-of-Experts (MoE)
• Total Parameters: 1T
• Activated Parameters: 32B
• Layers: 61 (including 1 dense layer)
• Attention Hidden Dimension: 7168
• MoE Hidden Dimension: 2048 (per expert)
• Attention Heads: 64
• Experts: 384
• Selected Experts per Token: 8
• Vocabulary Size: 160K
• Context Length: 128K
• Attention Mechanism: MLA
• Activation Function: SwiGLU

Evaluation Highlights

Kimi K2 outperforms competitors in various benchmarks:

• Coding: LiveCodeBench (53.7% Pass@1), SWE-bench Verified (71.6% Multiple Attempts Acc)
• Tool Use: Tau2 retail (70.6% Avg@4), AceBench (76.5% Acc)
• Math & STEM: AIME 2024 (69.6% Avg@64), MATH-500 (97.4% Acc)
• General Tasks: MMLU (89.5% EM), Livebench (76.4% Pass@1)

Deployment

• API: Available on Moonshot AI platform (OpenAI/Anthropic-compatible).
• Inference Engines: vLLM, SGLang, KTransformers, TensorRT-LLM.
• Model Format: Block-fp8 on Hugging Face.

Usage Examples

Chat Completion

def simple_chat(client, model_name):
    messages = [
        {"role": "system", "content": "You are Kimi..."},
        {"role": "user", "content": [{"type": "text", "text": "Self-intro"}]}
    ]
    response = client.chat.completions.create(
        model=model_name,
        messages=messages,
        temperature=0.6,
        max_tokens=256
    )
    print(response.choices[0].message.content)

Tool Calling

def tool_call_with_client(client, model_name):
    # Define tools and tool_map
    messages = [...]
    while finish_reason in [None, "tool_calls"]:
        completion = client.chat.completions.create(...)
        # Process tool calls and append results to messages

License

Modified MIT License.

Contact

[email protected]