Ornith-1.5-9B-MTPLX-4bit

A 4-bit affine quantization of ornith-ai/Ornith-1.5-9B for Apple Silicon, exported with MTPLX 2.12.2. The native MTP head and vision tower are included in BF16.

Quantization and files

Component	Format	File	Size
Language model	4-bit affine, group size 64; unquantized tensors in BF16	model.safetensors	5.038 GB
Native MTP head	BF16, one layer	mtp.safetensors	0.487 GB
Vision tower	BF16	model-vision.safetensors	0.912 GB

The complete download is approximately 6.46 GB (6.01 GiB). The tokenizer, chat template, processor configurations, weight index, and MTPLX runtime contract are included.

All 250 packed language-model weight matrices were checked to use 4 bits. The MTP head is preserved in BF16 with the Forge recipe mtp_policy: keep_bf16.

Usage

Install MTPLX following the official project instructions, then download this repository with the Hugging Face CLI:

hf download 0xzknw/Ornith-1.5-9B-MTPLX-4bit \
  --local-dir ./Ornith-1.5-9B-MTPLX-4bit

mtplx start cli --model ./Ornith-1.5-9B-MTPLX-4bit

For one completion using native MTP:

mtplx run --model ./Ornith-1.5-9B-MTPLX-4bit \
  --prompt "Write a Python function that reverses a string." \
  --max-tokens 512 --mtp --depth 3

The included runtime contract recommends the sustained profile and supports MTP depths up to 3. Its sampler defaults are temperature 0.6, top-p 0.95, and top-k 20. Run mtplx tune --model ./Ornith-1.5-9B-MTPLX-4bit --retune to measure performance on your Mac.

Provenance

• Source model: ornith-ai/Ornith-1.5-9B.
• Source revision: 489cb97981b8654bcfcf30ce1f94ed1b62e07b53.
• Exported with MTPLX 2.12.2 on October 4, 2026.
• This repository publishes the first local 4-bit export. The weight files and tokenizer are preserved unchanged from that export; publication metadata uses portable paths.

Validation and scope

The 1,275 tensor shapes and byte ranges were checked, including the 15 MTP tensors and 333 vision tensors. All 1,260 entries in the main weight index match the language and vision files. Local Forge verification covers autoregressive text generation and MTP depths 1, 2, and 3; the detailed runtime record is in mtplx_runtime.json.

Image and video inference were not separately tested for this publication. No independent downstream accuracy benchmark is claimed for this quantization; upstream benchmark scores describe the original model.

License

MIT, following the license declared by the upstream model card. Credit for the original model belongs to the Ornith team; MTPLX provides the conversion and inference tooling.