Kimi K2.5

Kimi K2.5 is Moonshot AI's flagship open-weight model — a 1.04 trillion parameter Mixture-of-Experts with 32B active parameters per token. It employs 384 experts with 8 activated per forward pass, using Multi-head Latent Attention (MLA) to cut memory bandwidth by 40-50%. Trained on 15 trillion mixed visual and text tokens, it delivers state-of-the-art coding (76.8% SWE-Bench Verified) and agentic capabilities with Agent Swarm technology coordinating up to 100 sub-agents. At 374 GB even at aggressive 2-bit quantization, Kimi K2.5 demands enterprise-grade hardware — multiple high-VRAM GPUs or a Mac with 400 GB+ unified memory. The native INT4 weights from Quantization-Aware Training make 4-bit quantization practically lossless compared to FP16. Available on Ollama with a cloud-backed tag for those without the local resources.