Ornith 1.5 → CMF — one file, one Rust binary, no Python

cargo install cortiq-cli
hf download infosave/Ornith-1.5-cmf ornith-1.5-9b-q4tp.cmf --local-dir .
cortiq run ornith-1.5-9b-q4tp.cmf --prompt "Explain quicksort in three sentences."

Ornith 1.5 is a Qwen3.5-family pair: a 9B dense hybrid and a 35B mixture with 3B active per token, both mostly linear-attention layers with full attention every fourth. These are the text towers converted to the CMF container — a single memory-mapped file read by cortiq, a Rust binary with no ML framework under it. GPU via Vulkan/Metal/DX12 with a CPU fallback.

file	params	layers	size
`ornith-1.5-9b-q4tp.cmf`	9B	32 (8 attention / 24 linear)	4.37 GB
`ornith-1.5-35b-a3b-q4tp.cmf`	35B total, 3B active	40 (10 attention / 30 linear), 256 experts, 8 per token + shared	17.4 GB

Both 4-bit tiled with ladder scales (q4tp), quantized straight from the bf16 checkpoints, MTP draft head included. The upstream repos are vision-language; these files carry the text tower — image input is not part of the container.

Speed

Steady-state decode, cortiq bench --core, single stream, A100 80GB over Vulkan, cortiq 0.5.99. The whole token runs as one submitted graph — linear-attention state, expert routing and the draft head included.

	A100 80GB (Vulkan)	Apple M4 (Metal)
9B	73 tok/s	20 tok/s
35B-A3B	65 tok/s (first token pays a one-off ~10 s expert upload)	—

The 9B on the M4 was verified against the published file itself — downloaded back from this repo, checksum matched, generation coherent.

For scale: this stack decodes the same class of files at ~55–65% of an RTX 5090's rate, so consumer-card numbers run higher.

Server and API

cortiq serve ornith-1.5-9b-q4tp.cmf --port 8080

Speaks the OpenAI API, so anything that talks to OpenAI talks to it.

Checksums

Each .cmf ships a .sha256 beside it.