Kimi K2.6 NVFP4 is NVIDIA's Blackwell-ready NVFP4 quantization of Moonshot AI's Kimi K2.6: a 1-trillion-parameter Mixture-of-Experts model with 32B active parameters per token, a 256K-token context window, thinking and instant modes, and weights built for long-horizon coding and agentic work.
Kimi K2.6 is a 61-layer Mixture-of-Experts transformer on the DeepSeek V3 architecture: 384 routed experts plus one shared expert, 8 experts selected per token, Multi-head Latent Attention, a 160K-token vocabulary, and a 256K-token context window. Moonshot trained it for long-horizon coding across languages such as Rust, Go, and Python, coding-driven front-end design, and agent swarms that scale to 300 sub-agents executing 4,000 coordinated steps. It runs in a thinking mode with interleaved reasoning and multi-step tool calls or an instant mode without reasoning, and a preserve-thinking option carries reasoning across turns for coding agents.
This checkpoint is NVIDIA's NVFP4 quantization of Kimi K2.6, produced with NVIDIA Model Optimizer by converting Moonshot's native INT4 weights to BF16 and then to NVFP4; only the linear operators inside the MoE transformer blocks are quantized. In NVIDIA's evaluation the NVFP4 checkpoint matches or exceeds the INT4 baseline on SciCode, MMMU Pro, AA-LCR, and IFBench and trails it by at most 0.5 points on GPQA Diamond and τ²-Bench Telecom. The quantized weights are governed by the NVIDIA Open Model License, and the underlying Kimi K2.6 weights by Moonshot AI's Modified MIT License. Full details are in Moonshot AI's announcement and the model card.
A Dedicated Endpoint is your own deployment of Kimi K2.6 NVFP4, on an inference stack that Modal tunes and autoscales. Bring fine-tuned weights if you have them. Compute is billed by the second, and only while it runs.
modal endpoint create --model nvidia/Kimi-K2.6-NVFP4