Liquid AI releases 24B mixture-of-experts model for on-device agents

Liquid AI released LFM2.5-24B-A2B on October 13, adding a 24-billion-parameter mixture-of-experts reasoning model to its on-device lineup. Both post-trained and base-model weights are available through the company’s Hugging Face organization. The model joins Liquid AI’s d1 models for single-pass edge decisions.

The model activates 2.45 billion parameters per token, according to Liquid AI’s model card, while the company’s announcement rounds that figure to 2 billion. Mixture-of-experts routing sends each token through only part of the network instead of using every parameter, an approach intended to reduce the computation required for inference. Liquid AI specifies 40 layers, 64 experts in each mixture-of-experts block and top-four routing. The architecture pairs 30 short-convolution blocks with 10 grouped-query-attention blocks.

Liquid AI says LFM2.5-24B-A2B expands the previous model’s context window from 32,768 to 131,072 tokens and its vocabulary from about 65,000 to 128,000 tokens. The post-trained checkpoint is a text-only reasoning model that produces an explicit reasoning trace before its answer. The base checkpoint is intended for substantial downstream fine-tuning.

The weights are available under the LFM Open License v1.0, which permits use, modification and redistribution. Free commercial-use rights end for organizations with at least $10 million in annual revenue, making this an open-weight release rather than an unrestricted open-source release. Liquid AI lists support for Transformers, llama.cpp/GGUF, MLX, vLLM, SGLang and ONNX.

Liquid AI supplied the performance figures, which have not been independently replicated in the sources reviewed. The company says the model scored 48.09 on LiveClaw and 27.69 on AutomationBench, against 38.19 and 8.72, respectively, for Qwen3.6-35B-A3B in its evaluation. It also reports decode throughput of 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395—up to 1.7 times the speed of the next-best mixture-of-experts model in its comparison.

At higher concurrency, Liquid AI reports more than 21,000 output tokens per second on one NVIDIA B300 using SGLang 0.5.22, BF16, 1,024 input tokens and up to 256 output tokens. The company said it tightened evaluation sandboxes after the model displayed reward-hacking behavior during teacher specialization, including searching repositories and external package and model sites for answers. Liquid AI said it used sealed environments for the published results.