infosave/LFM2.5-cmf

LFM2.5 → CMF — one file, one Rust binary, no Python

cargo install cortiq-cli
hf download infosave/LFM2.5-cmf lfm2.5-2.6b-q4tp.cmf --local-dir .
cortiq run lfm2.5-2.6b-q4tp.cmf --prompt "Explain what a Fourier transform does, in two sentences."

LiquidAI's LFM2.5 is a hybrid: most layers mix with a short gated convolution and only a few carry full attention, so the state a token needs is small and constant where the convolution runs. These are those checkpoints in 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; NVIDIA, AMD, Intel and Apple silicon read the same file.

file	params	layers	size
lfm2.5-230m-q4tp.cmf	0.23B	14 (5 attention / 9 conv)	132 MB
lfm2.5-2.6b-q4tp.cmf	2.70B	30 (8 attention / 22 conv)	1.43 GB
lfm2.5-8b-a1b-q4tp.cmf	8.3B total, 1B active	24 (6 attention / 18 conv), 32 experts, 4 per token	4.6 GB

All three are 4-bit tiled with ladder scales (q4tp), quantized straight from the bf16 checkpoints.

Speed

Steady-state decode, cortiq bench --core, single stream, cortiq 0.5.99+ (the whole-token graph learned this family's short-conv mixer and its sigmoid-routed MoE in 0.5.98/0.5.99 — earlier versions decode it an order of magnitude slower).

	A100 80GB (Vulkan)	Apple M4
230M	390 tok/s	138 tok/s
2.6B	141 tok/s	42 tok/s (CMF_GPU=0)
8B-A1B	124 tok/s	—

On a discrete card the whole token runs as one submitted graph, conv ring and expert routing included; the MoE's greedy output is token-identical to the CPU path. On Apple silicon the engine measures both arms at startup and picks; for the 2.6B the host arm wins there, and CMF_GPU=0 pins it.

Server and API

cortiq serve lfm2.5-2.6b-q4tp.cmf --port 8080

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

curl localhost:8080/v1/chat/completions -H 'content-type: application/json' \
  -d '{"model":"lfm2.5","messages":[{"role":"user","content":"Say hello"}]}'

Checksums

Each .cmf ships a .sha256 beside it.