Chronos-Ornith-1.5-35B-A3B — MLX 3-bit (text-only, fits 24 GB Macs)

A 3-bit MLX conversion of ornith-ai/Ornith-1.5-35B-A3B, made so the model runs on a 24 GB Apple Silicon Mac with room to spare.

Text-only. The vision tower is not included, which is what keeps this build at 15.2 GB (about 3 GB smaller than the multimodal 3-bit builds). If you need image input, use a multimodal build instead.

Conversion

Plain affine quantization with mlx_lm 0.31.3, with no calibration or mixed-precision recipe:

python -m mlx_lm convert --hf-path ornith-ai/Ornith-1.5-35B-A3B \
  -q --q-bits 3 --q-group-size 64

The result averages 3.5 bits per weight. mlx_lm keeps the MoE router gates at 8-bit.

Usage

pip install mlx-lm
mlx_lm.generate --model Lifeboat78/Chronos-Ornith-1.5-35B-A3B-MLX-3bit --prompt "Hello"

from mlx_lm import load, generate
model, tokenizer = load("Lifeboat78/Chronos-Ornith-1.5-35B-A3B-MLX-3bit")
print(generate(model, tokenizer, prompt="Hello", max_tokens=200))

Measured on an M5 MacBook Pro, 24 GB

These numbers come from a private 36-case suite for a domain-specific assistant covering routing, domain vocabulary, JSON output, reasoning, instruction following, false-premise handling and 32k-token retrieval. The comparison model is Unsloth's Qwen3.6-35B-A3B-UD-MLX-3bit, run with the same harness on the same day.

Known weakness: in one ungraded summarization case (a count broken into categories), this build miscounted the totals and gave inconsistent numbers, while the Qwen build got them right. Check its arithmetic on anything that matters. A plain 3-bit quantization without calibration may be part of the reason.

License and credit

The weights are under the MIT license from the base model. All credit for the model goes to ornith-ai. This repository only changes the quantization.