DeepReinforce debuts Ornith-1.0, a 397B MoE model

DeepReinforce has introduced Ornith-1.0-397B, the flagship entry in a new family of open language models. The release, published on Hugging Face, is a mixture-of-experts (MoE) system with 397 billion total parameters and is offered under the permissive MIT license.

The model is positioned for both general text generation and reasoning tasks, signaling that DeepReinforce intends Ornith to compete on the harder end of open-weight workloads rather than chat alone. As an MoE design, it routes inputs through a subset of specialized expert networks, an approach that lets a model carry a very large parameter count while keeping the compute used per token comparatively modest.

Why it matters

The choice of an MIT license is the most consequential detail for developers. Unlike the custom community licenses attached to many large open models, MIT places few restrictions on commercial use, redistribution, or fine-tuning. That makes a model of this scale unusually easy to build on.

• A 397B-parameter MoE flagship for a brand-new model family
• MIT licensing, friendly to commercial and derivative work
• Targeted at text generation and reasoning

As a debut release, Ornith-1.0 establishes a baseline for what DeepReinforce can ship. The questions that remain — context length, benchmark performance, and how the smaller siblings in the family will be sized — will determine whether Ornith earns a place alongside the established open MoE models. For now, the weights are available to anyone willing to provision the hardware to run them.