Liquid AI has launched LFM2.5-24B-A2B, an open-weight 24-billion parameter Mixture-of-Experts (MoE) model that activates approximately 2 billion parameters per token. Designed for on-device local agent loops, the text-only model was trained inside popular agent harnesses to execute multi-step tool use rather than standard chat. Base and post-trained weights have been released on Hugging Face under the LFM Open License v1.0, which remains free for commercial use until a company reaches $10 million in annual revenue.
Architecturally, LFM2.5-24B-A2B retains Liquid AI’s 40-layer hybrid setup, combining 30 double-gated short convolution blocks with 10 grouped query attention blocks. Key updates from its predecessor include an expanded context window of 131,072 tokens, a doubled vocabulary of 128,000 tokens, and pre-training expanded to 33 trillion tokens. On-device benchmarks show the model decoding 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395 using llama.cpp quantization.
In benchmark testing conducted by Liquid AI against competing models like Qwen3.6-35B-A3B and Gemma-4-26B-A4B-it, LFM2.5-24B-A2B led on agent task suites such as AutomationBench and LiveClaw. However, larger models retained advantages in knowledge and coding benchmarks like GPQA Diamond and SWE-Bench.
Why it matters
Enables fast, low-latency agentic workflows directly on consumer hardware like Apple M-series chips.
Hybrid convolution-attention architectures offer practical performance trade-offs for local execution.
Source: marktechpost.com



