New Maxwell AI Reads 256K Tokens, Runs Cheap
This AI can read a whole book at once—and it's cheap to run.
Maxwell-V1.0-Flash is a next-generation, high-performance Mixture-of-Experts language model developed by the Maxwell Team, with 35B total parameters and 3B active parameters per token (both excluding embeddings) and support for context lengths up to 256K tokens. It is designed for strong reasoning, agentic, and long-context capabilities at a practical inference cost.
Its features include Maxwell Block Sparse Attention, Manifold-Constrained Hyper-Connections, and a fine-grained Mixture-of-Experts backbone. Pre-training used approximately 26T tokens of multilingual data and the Muon optimizer on most hidden weight matrices; long-context training extends sequence length in stages from 4K to 256K. Post-training includes two-stage all-domain SFT, independent RL branches, and multi-teacher on-policy distillation.
- Maxwell-V1.0-Flash can read up to 256,000 words at once—like a whole book—and is designed to be cheap to run.
- It's especially good at reasoning and long tasks, and supports many languages, including some from Southeast Asia.
- It's a technical release for developers, not a ready-to-use app, so you won't use it directly yet.
Why It Matters
Future AI tools could read long documents cheaply, saving time and money for businesses and researchers.