技术规格

系统要求

不同量化方法和上下文大小的显存要求

Kimi K2 Thinking

Kimi K2 Thinking is a language model developed by Moonshot AI, engineered as a specialized thinking agent designed to perform complex, multi-step reasoning and dynamic tool invocation. The model is trained to interleave chain-of-thought processes with function calls, enabling it to execute intricate workflows such as autonomous research, coding, and writing that can persist over hundreds of sequential actions without coherence degradation. A key design principle is its native INT4 quantization, which is applied via Quantization-Aware Training (QAT) to achieve efficient inference, contributing to lossless reductions in inference latency and GPU memory utilization.

Architecturally, Kimi K2 Thinking operates on a sparse Mixture-of-Experts (MoE) paradigm, encompassing a total of 1 trillion parameters, with 32 billion parameters activated per inference pass. The model's internal structure includes 61 layers and employs a Multi-Head Latent Attention (MLA) mechanism with 64 attention heads. The activation function utilized is SwiGLU, and it features a vocabulary size of 160,000 tokens. It incorporates 384 experts, selecting 8 experts per token during processing, and is optimized for persistent step-by-step reasoning within its architectural constraints.

The model is characterized by a substantial 256,000-token context window, allowing for the processing of extensive textual inputs, which is particularly beneficial for long-horizon tasks, complex debugging, or comprehensive document analysis. This extended context, combined with its robust tool orchestration capabilities, enables Kimi K2 Thinking to maintain stable goal-directed behavior across 200 to 300 consecutive tool invocations. This capacity addresses a common limitation in prior models, which often exhibit performance degradation after a significantly fewer number of sequential steps.

没有可用的 Kimi K2 Thinking 评估基准。