Recovering most of what quantisation costs

Liquid AI released quantization-aware distillation checkpoints on 19 August for four models — LFM2.5 at 230M, 350M, 1.2B-Instruct and 2.6B — in Q4_0 GGUF. Rather than quantising a finished model, the lower-precision student is trained against the full-precision teacher, so the loss is absorbed during training instead of after it.

The recovery figures against the BF16 baseline: 97.1% at 230M, 96.5% at 350M, 97.4% at 1.2B and 96.6% at 2.6B.

The comparison that matters is the other one

Recovering accuracy against full precision is expected of any quantisation method. The useful claim is against the alternatives: these Q4_0 checkpoints match what Q5_K_M or Q4_K_M achieve, while decoding 3 to 33% faster. A smaller, cruder quantisation performing like a larger one is a straight saving in memory and time.

Evaluation covered six benchmarks — GPQA Diamond, MMLU-Pro, IFEval, IFBench, Multi-IF and BFCLv4 — plus maths tests scaled to model size, spanning reasoning, instruction following, tool use and agentic behaviour. Speed was measured on a MacBook Pro, a NucBox EVO-X2, a Galaxy S26 Ultra and a Raspberry Pi 5, across GPU and ARM CPU.

That hardware list is the point of the release. A 230M model recovering 97% of its accuracy at Q4_0 is a model that runs where there is no GPU at all, and the benchmarks were chosen to test whether it still follows instructions and calls tools once it gets there.

The announcement does not state the licence on the weights. For anything intended to ship inside someone else's product, that is the term to check on the model card first.