Language
Kimi K2 Thinking
#20moonshot-ai · thinking · seit 2025-11-06 · 8× · zuletzt 30. Juni 2026
17
Momentum
Kimi K2 Thinking is Moonshot AI's reasoning model released in November 2025, built on the Kimi K2 base (MoE, 1 trillion total parameters, 32 billion active per forward pass) and post-trained for multi-step reasoning with dynamic tool use. The model employs native INT4 Quantization-Aware Training (QAT), delivering approximately 2× faster inference and ~50% less GPU memory usage compared to FP16. It supports a 256k-token context window and is released as an open-weight model under a modified MIT license.
Historique du momentum
04.04.03.07.
Fonctionnalités
| Context Window (Tokens) | 256,000 tokens (256k); Max. output: 16,384 tokens (per Amazon Bedrock). Benchmarks were conducted at 256k context length. |