ezoushen/ornith-1.5-35b-a3b-mlx-4bit

Ornith 1.5 35B-A3B, MLX 4-bit with vision

Ornith-1.5-35B-A3B as an MLX affine 4-bit checkpoint that includes the vision tower and loads on Splash 1.1.0. We serve it with vision on our 1.1.0-based fork; on upstream 1.1.0 we have tested its language model only.

splash serve --model ezoushen/ornith-1.5-35b-a3b-mlx-4bit

The language model is ornith-ai/Ornith-1.5-35B-A3B-MLX-4bit, byte for byte. That repository has no vision tower, and Splash 1.1.0 stops on its config (reported upstream). This one adds both.

Independent redistribution, not endorsed by Ornith AI or Inco AI. MIT, as upstream.

What differs from ornith-ai's MLX 4-bit repo

Language model	unchanged: 1,757 tensors in four shards, affine 4-bit, group size 64 (router and shared-expert gate at 8 bits)
Vision tower	added: model-vision.safetensors, 333 BF16 tensors (0.89 GB), Ornith's tower, recovered from our Splash package by an exact inversion
config.json	adds rope_parameters.rope_type: "default" and a vision_config
Also added	preprocessor_config.json and a shard index that lists the vision tensors
chat_template.jinja	renders a system message after the first as a system turn instead of raising, the one-line change incoai's Qwen3.6 Splash package makes (since 2026-09-30)

Total size is 20.4 GB.

Which Ornith repo to use

• Just running Ornith on a Mac with Splash: ezoushen/ornith-1.5-35b-a3b-splash, the packed Splash package. It has the benchmarks and caveats.
• Serving Ornith beside a fine-tune of it: this repo. With the shared-expert-weights Splash fork, a fine-tune such as Umpire MLX 4-bit, served beside it from the same SPLASH_WEIGHT_CACHE, maps this model's prepared experts, embedding, head and draft instead of loading a second copy.

Licence

MIT, from Ornith AI (LICENSE). NOTICE lists every change.