Models · family

Liquid (LFM2)

4 MLX aliases · LiquidAI LFM2 / LFM2.5 · on-device / edge models · native MLX · lfm tool-call parser.

LiquidAI (the MIT spinout) builds LFM2 / LFM2.5 — a family of on-device, edge-oriented models designed to run natively on-device. rapid-mlx serves four MLX aliases spanning the range: a 24B-A2B MoE, an 8B-A1B MoE, and the 2.6B and 1.2B LFM2.5 models, all wired to the lfm tool-call parser. These are the low-RAM end of the catalog: the 1.2B fits comfortably on a laptop, and the sparse-MoE variants keep active-parameter cost small.

Pick one

rapid-mlx serve lfm2.5-1b-4bit
LFM2.5 1.2B · 0.66 GB download · the engine's fast pick for 8 and 16 GB Macs
rapid-mlx serve lfm2.5-2.6b-4bit
LFM2.5 2.6B · 1.6 GB download · the engine's smart pick for 8 GB Macs
rapid-mlx serve lfm2.5-8b-a1b-4bit
LFM2.5 8B-A1B MoE · 4.78 GB download
rapid-mlx serve lfm2-24b-a2b-4bit
LFM2 24B-A2B MoE · 13.42 GB download

family: Liquid
aliases: 4
vendor: LiquidAI (MIT spinout)
profile: on-device / edge · native MLX
install: rapid-mlx serve <alias>
OpenAI base URL: http://localhost:8000/v1

Download

Every alias on this page downloads with one command — the pull buttons in the tables below copy it. All 4 aliases on this page are mirrored on the rapid-mlx CDN — with automatic mid-pull fallback to Hugging Face if a mirror file slows down. Weights land in the standard Hugging Face cache, and rapid-mlx serve pulls automatically on first use. Live mirror status →

LFM2 / LFM2.5 · 4 aliases

On-device models from LiquidAI. All four use the lfm tool-call parser. Two of them — lfm2.5-2.6b-4bit and lfm2.5-8b-a1b-4bit — emit an implicit <think> channel and are routed through the qwen3 reasoning parser, so their reasoning stays in the reasoning field instead of leaking into content.

tool parser: lfm

alias	hf repo	tool parser	reasoning	flags	get it
lfm2-24b-a2b-4bit	lmstudio-community/LFM2-24B-A2B-MLX-4bit	lfm	—	moe	CDN
lfm2.5-8b-a1b-4bit	mlx-community/LFM2.5-8B-A1B-MLX-4bit	lfm	qwen3	moe	CDN
lfm2.5-2.6b-4bit	LiquidAI/LFM2.5-2.6B-MLX · `4bit/`	lfm	qwen3	dense	CDN
lfm2.5-1b-4bit	mlx-community/LFM2.5-1.2B-Instruct-4bit	lfm	—	dense	CDN

Pull & serve

rapid-mlx pull lfm2.5-1b-4bit

rapid-mlx serve lfm2.5-1b-4bit

Notes & caveats

On-device by design. LFM2 targets edge / laptop deployment — the 1.2B fits a small Mac; the A2B / A1B MoE variants keep active-parameter cost low.
Tool calling flows through the lfm parser; lfm2.5-2.6b-4bit and lfm2.5-8b-a1b-4bit also route their thinking through the qwen3 reasoning parser.

Where next

All families

The full family directory — pick another family or browse the hero deep dives.

Hunyuan (Hy3)

The other end of the range — Tencent's 295B MoE Ultra-class preview.

API surface

The OpenAI-compatible HTTP API: chat, completions, responses, embeddings, audio.