Ornith 1.5 arrives as a 397B MoE multimodal model

Ornith AI has published Ornith 1.5, a mixture-of-experts multimodal model with 397 billion total parameters, on Hugging Face. The model handles both images and text, positioning it as a vision-language system rather than a text-only chatbot.

The flagship isn't shipping alone. The 1.5 line also includes a compact 9B model and a 35B-A3B variant, the latter using an active-parameter design that keeps only a fraction of weights engaged per token. That spread lets teams pick a tier that fits their hardware budget while staying within the same family.

Why it matters

A few details stand out for people deciding whether to build on it:

• MIT license, which is unusually permissive for a model at this scale and allows commercial use with minimal restrictions.
• MoE architecture, meaning inference cost can stay lower than the headline 397B figure suggests.
• Multimodal by default, with image and text handling built in rather than bolted on.

Some specifics remain unlisted, including context length and detailed benchmark results, so the practical performance picture will firm up as the community tests the weights. For now, the combination of a large open MoE model and a genuinely open license makes Ornith 1.5 worth a look for anyone weighing self-hosted multimodal options.