Shypd.aiThe largest open index of AI toolsTutel is an optimized Mixture-of-Experts (MoE) library that supports models like GptOss, DeepSeek, Kimi-K2, and Qwen3. It offers a parallel solution for training and inference with dynamic behaviors, supporting FP8, NVFP4, and MXFP4 precisions.
Trending
Also listed in
About
Tutel is an optimized Mixture-of-Experts (MoE) library developed by Microsoft, designed to enhance the performance of large language models. It provides a parallel solution for both training and inference, accommodating dynamic behaviors often found in modern AI workloads. Tutel supports various advanced precision formats including FP8, NVFP4, and MXFP4, making it compatible with high-performance GPUs like A100, H100, B200, and MI300. The library is built on PyTorch (recommended version >= 2.0) and supports CUDA and ROCm for GPU acceleration, as well as standard CPU operations. It includes optimized gating APIs for models like Qwen3 and DeepSeek, and offers features like 'No-penalty Parallism/Sparsity/Capacity/.. Switching' for efficient resource utilization.
Capabilities
Pricing & Plans
Open Source
Free
FAQs
Trending
