Mistral's Leanstral 1.5 puts 119B in a lean MoE

Mistral AI has published Leanstral 1.5, a mixture-of-experts language model with 119 billion total parameters but only about 6 billion active for any given token. The release is available on Hugging Face under the permissive Apache-2.0 license, which allows commercial use and modification without the usage restrictions attached to many competing open-weight models.

The naming hints at the design goal. A sparse MoE routes each token through a small subset of specialized experts, so a model can hold a large pool of knowledge while keeping the compute per forward pass closer to that of a much smaller dense model. In Leanstral's case, the roughly 6B active footprint is what determines inference cost and latency, not the full 119B.

Why it matters

The economics of running large models remain the biggest barrier to open deployment, and sparse activation is one of the more practical answers.

• Total capacity: 119B parameters across experts
• Active per token: about 6B, keeping serving costs down
• License: Apache-2.0, friendly to commercial and downstream use
• Focus: text generation with reasoning among its listed modalities

Mistral has built much of its reputation on efficient, openly licensed models, and Leanstral 1.5 continues that pattern by targeting teams that want a capable model they can host themselves without heavy hardware demands. Full context-window details and benchmark figures were not specified in the release record; readers should consult the model card for configuration and evaluation specifics as they are published.