Liquid AI releases open-weight LFM2-24B-A2B

Liquid AI announced LFM2-24B-A2B, the largest model in its LFM2 family. It is a sparse mixture-of-experts (MoE) model with 24B total parameters and 2.3B active per forward pass. The open weights are on Hugging Face. The company says the model fits in 32GB of RAM, so it can run on laptops and desktops with integrated GPUs or NPUs.

The model uses the LFM2 hybrid backbone, which mixes gated short-convolution blocks with grouped query attention at a ratio of about 3:1. Compared with LFM2-8B-A1B, it grows from 24 to 40 layers and from 32 to 64 experts per MoE block, while keeping top-4 routing and a 2048 hidden dimension. Total parameters triple, but active parameters grow only about 1.5x.

Liquid says average scores on GPQA Diamond, MMLU-Pro, IFEval, IFBench, GSM8K and MATH-500 rise log-linearly from LFM2-350M up to the 24B model. The release is an instruct model without reasoning traces and has had only light post-training.

On a single H100 SXM5 running vLLM, the company reports about 26.8K total tokens per second at 1,024 concurrent requests. It says that beats gpt-oss-20b and Qwen3-30B-A3B-Instruct-2507. Liquid also ran llama.cpp comparisons on an AMD Ryzen AI Max+ 395, but the numbers for those tests were not in the text provided.

In a partner launch post, Liquid gave figures for smaller devices. The model decodes at 35.4 tokens per second on a Samsung Galaxy S25 Ultra (Snapdragon 8 Elite, Hexagon NPU) using Nexa AI's SDK. On an Apple M4 Pro CPU with INT8 quantization, the Cactus engine reaches 229 tokens per second prefill and 27 tokens per second decode. In the cloud, Together AI offers the model as serverless inference with a 99.9% reliability SLA, under the model id liquid-ai/lfm2-24b-a2b.

The model works on day one with llama.cpp, vLLM and SGLang, with GGUF quantizations from Q4_0 to F16. Pre-training is still running and has used 17T tokens so far. Liquid plans an LFM2.5-24B-A2B with more post-training and reinforcement learning once pre-training finishes.