Duckets/Ornith-1.5-122B-MLX-4bit

Ornith-1.5-122B — MLX 4-bit

An MLX 4-bit conversion of Ornith-1.5-122B by Ornith AI (ornith-ai) — all credit for the model itself, its training, and its capabilities belongs to them. This repo is a downstream conversion built so the model runs natively on Apple Silicon with mlx-vlm / mlx-lm and in LM Studio.

Ornith-1.5-122B is the flagship of the Ornith-1.5 family: a 122B-parameter MoE model with strong instruction following, task decomposition, coding, reasoning, and agentic benchmark results, released open-weight under the MIT license. See the upstream model card for the full description and benchmarks.

Conversion details

Source	ornith-ai/Ornith-1.5-122B (BF16)
Quantization	MLX affine 4-bit, group size 64, 4.544 bits per weight (router/shared-expert gates kept at 8-bit by the converter, as recorded in config.json)
Converter	mlx_vlm.convert (mlx-vlm), run on CPU after a GPU attempt hit a Metal command-buffer timeout during save
Size on disk	64.86 GiB, 14 safetensors shards
Total parameters	122B (MoE)
Architecture	Qwen3_5MoeForConditionalGeneration (qwen3_5_moe), 48 layers, 256 experts, top-8 per token
Context length	262,144 tokens
Inputs	Text and images (vision tower carried over from the source)
Output	Text

Verified

• Smoke test on a 128 GB Apple Silicon Mac Studio (2026-10-20), mlx_lm.generate, prompt "What are you, and what can you do?", 64 tokens: coherent reasoning and answer text, generation at 68.7 tokens/sec, peak memory 68.9 GB.
• Vision verified in LM Studio (2026-10-20): shown a test image, the model read a random 6-character code from it exactly and correctly described the shapes, colors, and positions in the image.
• Convert completed from the full BF16 source (62 source shards, 233 GB) with the quantizer reporting 4.544 bits per weight.

Usage

With mlx-lm:

pip install -U mlx-lm

python -m mlx_lm.generate \
  --model Duckets/Ornith-1.5-122B-MLX-4bit \
  --prompt "What are you, and what can you do?" \
  --max-tokens 256

With mlx-vlm (for image input):

pip install -U mlx-vlm

python -m mlx_vlm.generate \
  --model Duckets/Ornith-1.5-122B-MLX-4bit \
  --prompt "Describe this image." \
  --image /path/to/image.png

With LM Studio: download this repo into your LM Studio models folder and load it from the MLX models list. Note the model needs roughly 69 GB of memory at load, so on a 128 GB machine leave headroom alongside anything else loaded.

Limitations

• 4-bit quantization can differ slightly from the BF16 original on borderline outputs; use the upstream BF16 release for full-precision evaluation.
• Beyond text and image input (verified above), other upstream capabilities are carried over unchanged but were not separately re-verified in this port.

License

MIT — the same license as the upstream release. See the upstream LICENSE. This is a derivative conversion of Ornith AI's model; all credit for the model itself belongs to them.

---

Conversion and packaging by Juno for Duckets, 2026-10-20.