Liquid AI LFM2.5 Q4_0 Recovers 97% Accuracy via Distillation

On August 19, 2026, Liquid AI announced new Q4_0 GGUF checkpoints for its LFM2.5 model family. The release utilizes a novel Quantization-Aware Distillation (QAD) technique to address the accuracy degradation typically seen when compressing high-precision models for edge hardware. By distilling a high-precision teacher model directly into a quantized student model, the process allows the LFM2.5 series to recover up to 97% of the accuracy lost during 4-bit quantization.

This release targets developers building agentic workloads and on-device applications. The Q4_0 GGUF format is heavily optimized for edge runtimes like llama.cpp. By using QAD, the student model learns to operate effectively within the mathematical constraints of 4-bit precision from the start, avoiding the performance penalties of applying quantization strictly as a post-training step.

Quantization-Aware Distillation Metrics

The QAD process produces checkpoints that match the memory footprint and throughput of native Q4_0 GGUF files while yielding output quality comparable to larger Q5_K_M formats.

Model	BF16 Accuracy Retained
LFM2.5-230M	97.1%
LFM2.5-350M	96.5%
LFM2.5-1.2B-Instruct	97.4%
LFM2.5-2.6B	96.6%

The LFM2.5-2.6B variant features 2.69B parameters spread across 30 layers. The architecture utilizes 22 double-gated short convolution blocks and 8 GQA layers to balance parameter efficiency with reasoning capacity.

Hardware Throughput

Liquid AI tested these models across diverse hardware platforms, targeting developers running local models on edge architectures. The testing matrix included Apple M5 Max and NucBox EVO-X2 for GPU benchmarks, alongside Samsung Galaxy S26 Ultra and Raspberry Pi 5 for Arm CPU inference.

On an Apple M5 Max, LFM2.5-2.6B reaches decode speeds of 220 tok/s. When running on an AMD Ryzen CPU, the same model hits 113 tok/s. The entire process executes within a memory footprint of under 2.5 GB. This memory efficiency dictates how you allocate system resources for concurrent agent routines on consumer hardware.

Evaluation Suite

The evaluation phase for the QAD checkpoints utilized a strict benchmark suite encompassing GPQA Diamond, MMLU-Pro, IFEval, IFBench, Multi-IF, and BFCLv4. Liquid AI applied scale-specific math evaluations to validate the models. GSM8K measured reasoning in the smaller parameter configurations, while AIME25 challenged the 1.2B and 2.6B variants.

This rollout concludes Liquid AI’s August 2026 LFM2.5 series deployment schedule, which previously introduced the LFM2.5-VL-3B vision-language model on August 12 and the LFM2.5-Embedding-350M on August 18.

For engineers optimizing edge AI pipelines, understanding quantization in AI is no longer just about selecting a post-training script. If you build inference runtimes for constrained devices, replacing standard Q4_0 formats with QAD-distilled checkpoints reclaims significant base model accuracy without modifying your existing llama.cpp integration.