maxwell-model/Maxwell-V1-Flash: A 35B Mixture-of-Experts Language Model

huggingface.co4 min read

The Maxwell Team released Maxwell-V1.0-Flash, a next-generation 35B Mixture-of-Experts language model featuring 3 active parameters per token and support for up to 256K context lengths. The model employs Maxwell Block Sparse Attention and Manifold-Constrained Hyper-Connections to balance strong reasoning, agentic, and long-context capabilities at practical inference costs. Open article →

More on this topic