Liquid AI releases on-device MoE LFM2-8B-A1B
Liquid AI has released an early checkpoint of LFM2-8B-A1B, a sparse model meant to run on phones and laptops with the quality of a larger model and the compute cost of a 1.5B-parameter one.
Liquid AI announced an early checkpoint of LFM2-8B-A1B. It is the company's first on-device Mixture-of-Experts (MoE) model, with 8.3B total parameters and 1.5B active per token. Liquid says the model needs more memory than a dense model but delivers better quality and speed, and that its quantized versions fit on high-end phones, tablets and laptops.
The model keeps the LFM2 backbone of 18 gated short-convolution blocks and 6 grouped-query attention blocks. Every layer except the first two, which stay dense for stability, has an MoE block. Each block holds 32 experts and sends every token to 4 of them. The router uses normalized sigmoid gating with adaptive routing biases. Liquid presents the model as a challenge to the common view that MoE doesn't work well at small sizes.
The base model was pre-trained on about 12T tokens: roughly 55% English, 25% multilingual text and 20% code. Training is ongoing. On Liquid's internal evaluations, the model scored 64.84 on MMLU, 37.42 on MMLU-Pro (11.46 points above LFM2-2.6B), 84.38 on GSM8K and 74.2 on MATH500. Results against Llama-3.2-3B-Instruct, SmolLM3-3B and gemma-3-4b-it were mixed. Qwen3-4B-Instruct-2507 led on most of the benchmarks.
The model runs on llama.cpp and ExecuTorch, with vLLM support "coming soon." Liquid says it decodes significantly faster than models with similar active parameter counts, such as Qwen3-1.7B, on a Samsung Galaxy S24 Ultra and an AMD Ryzen HX370 using Q4_0 quantization. The signal does not include the throughput numbers, the coding scores or licensing terms.