Ornith-1.5-122B Targets 48GB GPUs

This is exactly the kind of community optimization that makes giant open models useful beyond datacenter hardware.

• –The base model is designed for an 80GB GPU; fitting it into 48GB could significantly expand local experimentation.
• –Quantization quality, context limits, and generation speed still need independent validation—the official benchmarks cover the original model, not this compressed build.
• –Ornith-1.5-122B combines a Qwen3.5-MoE architecture with reasoning, tool calling, and coding-agent support.
• –Its 262K context window remains memory-intensive, so weight compression alone may not make full-context workloads practical.
• –The broader Ornith-1.5 family uses self-generated tasks, scaffolds, and rollouts for reinforcement learning, an interesting alternative to static training curricula. [Ornith blog](https://ornith.ai/ornith_1_5.html) [EXL3 quantization details](https://huggingface.co/Mia-AiLab/Ornith-1.5-122B-EXL3-EAQv1-4-bpw)