back to local model checker

LFM2 24B

MoE

Liquid AI · 24B (2.3B active) · Mixture of Experts

Hybrid MoE with convolution+attention layers — 2.3B active

Liquid AI
Downloads

16.5K

Likes

275

Released

2025-11

Context

32K tokens

Use cases

chatedgerag
Total experts64
Active experts4
Active params2.3B

Your hardware

16 GB Reference Desktop

Quant
Q4_K_M
Memory
12.8 GB
Speed
~16 tok/s
Fit
Tight fit

Quantization options

What is quantization?
QuantBitsVRAMDiskQualityStatus
Q2_K28.2 GB7.3 GBlowDecent · ~25 t/s
Q3_K_M311.3 GB10.3 GBmoderateTight fit · ~18 t/s
Q4_K_M412.8 GB11.7 GBgoodTight fit · ~16 t/s
Q5_K_M515.9 GB14.7 GBgoodBarely runs · ~13 t/s
Q6_K618.9 GB17.6 GBexcellentBarely runs · ~5 t/s
Q8_0825.1 GB23.5 GBexcellentBarely runs · ~3 t/s
F161649.7 GB46.9 GBlosslessToo heavy · ~1 t/s

About this model

LFM2 24B is a 24B sparse mixture-of-experts (MoE) model activating 2.3B active per token from Liquid AI in the LFM family, designed primarily for chat, edge, rag workflows. Supports up to 32K context tokens; start at 4K–8K with Q8_0 KV cache on tight GPUs.

Hardware planning baselines start at 13.4 GB minimum memory and 22.4 GB recommended memory. At Q4_K_M, weights plus baseline runtime overhead use roughly 12.8 GB of RAM/VRAM and 11.7 GB of disk space. Because MoE routing evaluates only a subset of experts on each generated token, LFM2 24B delivers faster decoding throughput per gigabyte of memory bandwidth than a dense checkpoint of identical total size—provided the full 12.8 GB checkpoint stays resident without SSD paging.

Deploy LFM2 24B locally via Ollama (`lfm2:24b`), LM Studio, or Ollama with Flash Attention enabled. Under Liquid AI Open Weights License, review official weight cards before commercial redistribution.

Highlights

  • Local Fit Baseline13.4 GB minimum and 22.4 GB recommended RAM/VRAM.
  • Primary QuantQ4_K_M targets 12.8 GB memory and 11.7 GB disk.
  • Context & Capabilities32K max context · Sparse MoE.

System requirements by device

Memory planning targets 13.4 GB minimum RAM/VRAM and 22.4 GB recommended memory. The baseline quant is Q4_K_M, estimated at 12.8 GB memory and 11.7 GB disk. Actual runtime usage shifts with context length, KV cache precision, and OS background overhead.

DeviceNotes
Mac (Apple silicon)Use unified memory headroom (~75% physical RAM max) as the primary fit signal. Start with Q4_K_M and close heavy browser tabs first.
Windows / Linux GPUKeep weights plus KV cache inside dedicated VRAM so Windows WDDM does not spill layers over PCIe into Shared GPU Memory.
Phones & Compact Edge BoardsNot practical on phones (13.4 GB min RAM exceeds Android LMK and iOS Jetsam limits). Pick a 2B–4B model on mobile.

Start with a short session and leave memory for the operating system. A file fitting on disk does not guarantee that it will load in memory.

Estimated memory by quantization

These are rough estimates, not tested minimum requirements. Check the actual file and runtime documentation; context, offloading and operating-system overhead can change the memory needed.

Quant TierEstimated memoryDisk SizeQuality
Q2_K8.2 GB7.3 GBlow
Q3_K_M11.3 GB10.3 GBmoderate
Q4_K_M12.8 GB11.7 GBgood
Q5_K_M15.9 GB14.7 GBgood
Q6_K18.9 GB17.6 GBexcellent
Q8_025.1 GB23.5 GBexcellent
F1649.7 GB46.9 GBlossless

How to run LFM2 24B on your device

Start with a short prompt and the smallest quant that still gives the quality you need. A clean first run is more useful than loading a larger file with no memory headroom.

Mac

Run LFM2 24B on Mac

  1. Install Ollama or LM Studio on macOS Apple Silicon.
  2. Pull Q4_K_M (12.8 GB) first and confirm it stays below ~75% of your Mac's unified memory.
  3. Enable Flash Attention and cap initial context at 8K until memory pressure stays below 22.4 GB.
  4. Consult the Mac guide for Activity Monitor tuning and Metal offload tips.
Device how-to guide
Windows

Run LFM2 24B on Windows

  1. Install LM Studio for Windows for visual VRAM inspection, or Ollama when you want a localhost API endpoint.
  2. Load Q4_K_M (~12.8 GB) and verify Dedicated GPU Memory stays within VRAM without spilling into Shared GPU Memory.
  3. Reserve at least 13.4 GB memory minimum and target 22.4 GB for comfortable context growth.
  4. Enable Q8_0 KV cache or drop one quant tier if generation slows down over PCIe.
Device how-to guide

LFM2 24B vs alternatives

ModelParamsContextMin RAMLicense
LFM2 24B24B (2.3B active)32K13.4 GBLiquid AI
Devstral Small 2 24B24B256K13.4 GBApache 2.0
Mistral Small 3.1 24B24B128K13.4 GBApache 2.0
DiffusionGemma 26B-A4B IT26B (4B active)256K14.5 GBApache 2.0
GPT-OSS 20B21B (3.6B active)128K11.7 GBApache 2.0
Gemma 2 27B27B8K15.1 GBGemma

Use this table to choose the smallest nearby model that still matches the job. A related model with lower RAM can be the better local choice when the current page fits only at the edge of available memory.

Frequently asked questions

What is LFM2 24B?
LFM2 24B is a 24B sparse mixture-of-experts model (2.3B active active per token) from Liquid AI, supporting up to 32K tokens with 13.4 GB minimum RAM and 22.4 GB recommended RAM.
How much VRAM or RAM does LFM2 24B require?
At Q4_K_M, LFM2 24B occupies roughly 12.8 GB of active memory and 11.7 GB on disk. Add 2–6 GB of spare headroom for KV cache buffers, runtime scratch space, and OS apps (22.4 GB recommended).
What is the best way to run LFM2 24B locally?
Use `ollama run lfm2:24b` for a quick CLI/API setup, or load Q4_K_M in LM Studio or llama.cpp with Flash Attention enabled when you want explicit KV cache and layer control.
Can LFM2 24B run on Android or iPhone?
LFM2 24B requires 13.4 GB minimum memory, placing it outside practical 8 GB smartphone limits. Run it on a 16 GB+ Mac or desktop GPU instead, and pick a 2B–4B checkpoint for phones.
How does LFM2 24B compare to Devstral Small 2 24B?
Devstral Small 2 24B (24B, 13.4 GB min RAM) sits closest in size. Compare active parameters, context ceiling (32K vs 256K), and license terms (Liquid AI) before switching.