Basalt Labs drops 1.57T MoE Monolith-1.0
Basalt Labs has released Monolith-1.0, an open-weight 1.57-trillion-parameter Mixture-of-Experts reasoning model under the MIT license. Trained on 60 trillion tokens, the model supports a native 1-million-token context window and integrates grouped-query attention, fine-grained routing, and multi-token prediction heads.
Releasing a trillion-parameter-class reasoning model under an MIT license dramatically lowers the barrier to entry for advanced open-weights AI, though the massive hardware requirements to host it will keep it out of reach for average developers.
- –Massive 1.57T parameters (49.5B active per token) makes it one of the largest open MoE models to date.
- –Supports a native context length of 1,048,576 tokens via a two-stage YaRN schedule.
- –MIT License enables fully permissive commercial adoption, modification, and redistribution.
- –Running the model requires enterprise-grade hardware clusters (e.g., GB300 NVL72 racks or CloudMatrix-384 super-pods).
DISCOVERED
2h ago
2026-07-18
PUBLISHED
2h ago
2026-07-18
RELEVANCE
AUTHOR
FahadPrimeX