Ornith-1.5-35B-A3B-OptiQ-4bit

mlx-community/Ornith-1.5-35B-A3B-OptiQ-4bit#

Built with mlx-optiq, the MLX-native toolkit to quantize, fine-tune, and serve LLMs locally on Apple Silicon, no PyTorch and no cloud. All OptiQ quants · Docs

OptiQ mixed-precision quant of ornith-ai/Ornith-1.5-35B-A3B, a 35B mixture-of-experts vision-language model with 3B active parameters and a bundled MTP speculation head. 24 GB on disk.

What it is#

Following the naming llama.cpp uses for its mixed quants, the "4bit" label denotes the family, not the weighted average.

Run it#

The MoE arch and the MTP/vision sidecars register through OptiQ, so import optiq once before loading:

pip install "mlx-optiq>=0.4.27"

import optiq  # registers the arch + MTP/vision sidecars
from mlx_lm import load, generate

model, tok = load("mlx-community/Ornith-1.5-35B-A3B-OptiQ-4bit")
prompt = tok.apply_chat_template(
    [{"role": "user", "content": "Explain mixture-of-experts routing in two sentences."}],
    tokenize=False, add_generation_prompt=True,
)
print(generate(model, tok, prompt=prompt, max_tokens=400))

For image input plus an OpenAI- and Anthropic-compatible endpoint with mixed-precision KV cache:

optiq serve --model mlx-community/Ornith-1.5-35B-A3B-OptiQ-4bit

This is a reasoning model, so give it a generous token budget.

Using the head outside optiq serve: optiq/mtp.safetensors holds Qwen's original MTP tensors, so its seven RMSNorm weights (mtp.*norm*) are zero-centred, applied as 1 + w. The main weights have the +1 folded in, as MLX expects. Add 1.0 to all seven norms when loading the head into another runtime.

Links#

• Project website: mlx-optiq.com
• All OptiQ quants: mlx-optiq.com/models
• Base model: ornith-ai/Ornith-1.5-35B-A3B