Ornith-1.5-122B: a new 122B MOE model for text-generation on HuggingFace

On October 16, 2026, the model ornith-ai/Ornith-1.5-122B with the text-generation task was published on HuggingFace. The author describes it as a member of the Ornith-1.5 family and claims it runs on a single DGX, as well as the ability to run on mobile devices via the quantized variant Ornith-1.5-122B-Mobile.

What it means

The model contains 125.1 billion parameters, uses the MOE architecture and is distributed under the MIT license. The full weights repository in safetensors format takes up 233 GB, the largest file is 4.7 GB. By default it runs via transformers. The site already has 26 quantized builds in GGUF, EXL, NVFP4, GPTQ and MLX formats. The smallest working build — IQ4_XS from AtomicChat/Ornith-1.5-122B-GGUF — weighs 57.9 GB. Running such a build locally will require a corresponding amount of system RAM or VRAM, depending on the loading method. More news about local AI tools for 3D and video — in the CPU3D news feed.