Liquid AI · 2026-08-19 · notable
LFM2.5 Q4_0 — Liquid AI's 4-bit models keep 96%+ of full accuracy
Liquid AI released Q4_0 GGUF checkpoints for LFM2.5 at 230M, 350M, 1.2B and 2.6B. Quantization-aware distillation keeps 96.5-97.4% of the BF16 benchmark average and decodes 3-33% faster on phones, mini-PCs and a Raspberry Pi 5.
LFM2.5's new Q4_0 checkpoints are trained as 4-bit models rather than squeezed into 4 bits afterwards.
Key specs
| Accuracy kept vs bf16 | 96.5-97.4% |
|---|---|
| Sizes released | 230M, 350M, 1.2B, 2.6B |
What is it?
LFM2.5 Q4_0 is a set of four on-device checkpoints - 230M, 350M, 1.2B Instruct and 2.6B - that Liquid AI published on August 19, 2026. Each ships as a GGUF file in Q4_0, the simplest and smallest 4-bit layout. What is different is how they were made: by training, not by compressing a finished model.
How does it work?
Quantization-Aware Distillation is the technique behind the gain. A high-precision teacher model trains a student whose weights are already quantized, so the student learns to work inside 4-bit limits instead of being squeezed into them at the end. Liquid AI reports the resulting checkpoints hold 96.5% to 97.4% of the BF16 model's average benchmark score, depending on size.
Why does it matter?
Q4_0 is normally the format people avoid, because it is the fastest but loses the most quality, pushing them toward larger K-quants. Liquid AI measured these builds matching Q5_K_M quality at 230M and 350M and Q4_K_M quality at 1.2B and 2.6B, while decoding 3-33% faster on a MacBook Pro, a NucBox EVO-X2, a Galaxy S26 Ultra and a Raspberry Pi 5.
Who is it for?
on-device and edge developers
Try it
LiquidAI/LFM2.5-1.2B-Instruct-GGUF