What makes Liquid AI's new 24B model run on just 2B active parameters?

How the architecture keeps compute low

The model uses 40 layers split between two attention types. Thirty layers rely on gated short convolutions. Ten layers use grouped-query attention.

Each mixture-of-experts block contains 64 experts. Only four experts activate for any given token. This top-4 routing means the active parameter count stays at 2B while the total reaches 24B.

The vocabulary spans 128K tokens. Supported languages include English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, and Polish.

Reasoning that shows its work

Post-training inside agent harnesses made tool calling reliable. Unlike its predecessor LFM2-24B-A2B, this version always thinks before answering. The chat template opens every assistant turn with a <reasoning> tag. Responses contain an explicit chain of thought before the final answer. You must leave enough room in your max token budget for the reasoning trace.

Generation defaults sit at temperature 0.1, top_k 50, and repetition_penalty 1.05.

Running it locally today

Transformers support requires version 5.0.0 or newer. The HF repo ID is LiquidAI/LFM2.5-24B-A2B. Load it in bfloat16 with device_map set to auto. Flash Attention 2 works on compatible GPUs.

For MLX you need mlx-lm 0.32.0 or later. Earlier versions route experts incorrectly. Quantized builds ship as MLX-4bit, MLX-5bit, MLX-6bit, and MLX-8bit. GGUF, ONNX, vLLM, and SGLang paths are also documented.

What the documentation leaves out

Benchmark scores against other models do not appear on the page. Power draw measurements for on-device inference are absent. The license terms link to a separate user guide rather than stating them inline. How many devices currently run this model in production is not disclosed.