A GGUF conversion of LiquidAI's LFM2.5-1.2B-Thinking for llama.cpp / vllm.cpp, with a mixed quantization: the bulk of the weights are 4-bit (Q4_0) while the most sensitive tensors keep a higher precision. This variant is one of the four models benchmarked in the three-engine report:

This is a quantization of LFM2.5-1.2B-Thinking by Liquid AI. We are deeply grateful to Liquid AI for the LFM2.5 family, its gated-delta / shortconv architecture, and for publishing the weights openly. This work builds directly on theirs, and we thank them profusely. See the report above for the full acknowledgement.