LiquidAI-LFM2.5-230M-GGUF

LFM2.5-230M is Liquid AI's most compact hybrid model to date, a 230-million-parameter, general-purpose instruction-tuned text model built on the LFM2 architecture with extended pre-training (19T tokens) and reinforcement learning, designed specifically for on-device deployment in the tightest memory and compute budgets. Its 14-layer architecture combines 8 double-gated LIV convolution blocks with 6 GQA blocks, supports a 32,768-token context window across 10 languages, and was distilled from the larger LFM2.5-350M before being refined with multi-stage reinforcement learning, making it well-suited for agentic tasks like tool use and data extraction rather than reasoning-heavy workloads such as advanced math, code generation, or creative writing. It delivers strong edge inference throughput — 213 tok/s decode speed on a Galaxy S25 Ultra and 42 tok/s on a Raspberry Pi 5 — and despite its tiny size, outperforms similarly-scaled competitors like Granite 4.0-350M and LFM2-350M on benchmarks including IFEval (71.71), BFCLv3 (43.26), and Multi-IF (37.70), while supporting native function calling via Pythonic tool calls and ChatML-style chat templates, with deployment options spanning Transformers, vLLM, llama.cpp (GGUF), ONNX, and MLX formats.

Model Files

llama.cpp