Liquid AI Releases 24B LFM2.5-24B-A2B Model for Local Agents

On HuggingFace, on October 12, 2026, the model LiquidAI/LFM2.5-24B-A2B was published with the text-generation task. This is the largest and final model in the LFM2.5 family: 24 billion parameters, of which only 2 billion are active per token. The author positions it as a model for fast private on-device agents and compares it with Qwen3.6-35B-A3B, which has roughly 50% more both total and active parameters.

What it means

The full weights in safetensors take up 44.7 GB, the largest file is 4.6 GB. The model runs via transformers, the architecture is MoE. The platform already has 28 quantized builds: GGUF, FP8, MLX, GPTQ and NVFP4. The smallest working build is Q4_0 at 12.6 GB; we don't count builds below 4 bits. License — LFM Open License v1.0: commercial use is permitted up to annual revenue of 10 million US dollars; for non-profit organizations the threshold does not apply. When distributing, you must include the license, note changes and retain attribution notices. For local running, 12.6 GB in Q4_0 is a level that fits into the VRAM of modern consumer GPUs, and on Apple Silicon the model runs via MLX. The author claims 184 tokens per second on an Apple M5 Max, but on more modest hardware the speed will be lower. More on local AI tool news — in the CPU3D news feed.