Liquid AI released LFM2.5-8B-A1B, an edge model optimized for tool calling on consumer hardware. The update expands the context window from 32K to 128K tokens and increases pretraining data from 12T to 38T tokens. A doubled vocabulary (65K to 128K) improves tokenization efficiency for non-Latin languages like Hindi, Thai, and Arabic.

Unlike its predecessor, LFM2.5-8B-A1B operates as a reasoning-only model, generating explicit chain-of-thought before final answers. This approach leverages Mixture-of-Experts architecture where fewer active parameters make each reasoning token computationally cheap. The model uses large-scale reinforcement learning and targeted preference optimization to reduce hallucinations and doom loops in long reasoning traces.

Benchmark gains are substantial across knowledge, math, and agentic tasks. The AA-Omniscience Index improved from -78.42 to -24.70, while non-hallucination rates jumped 56 points to 63.47. MATH500 scores rose to 88.76, and the model competes with much larger dense models on instruction following benchmarks.

Both base and post-trained variants are available today on Hugging Face and the Liquid AI Playground. The model supports llama.cpp, MLX, vLLM, and SGLang for local deployment, making high-quality AI reasoning accessible on entry-level laptops without cloud dependencies.