What changes when a 4-bit model keeps 97 percent of its full-precision accuracy?

Liquid AI has released new 4-bit checkpoints for four LFM2.5 models. They were trained with a method called Quantization-Aware Distillation and recover most of the accuracy usually lost when shrinking a model to four bits.

How the checkpoints are built

A high-precision teacher model is distilled into a quantized student model. The student starts quantized and learns directly from the teacher during training. This differs from post-training quantization where a full-precision model is compressed after the fact. The result keeps the low memory footprint and high throughput of native Q4_0 GGUFs. Recovery reaches 97 percent of the BF16 average accuracy lost to quantization. Quality returns.

What the benchmarks show

Four models were tested: LFM2.5-230M, LFM2.5-350M, LFM2.5-1.2B-Instruct, and LFM2.5-2.6B. The benchmark suite spans reasoning, instruction-following, tool use, and agentic capabilities. GPQA Diamond, MMLU-Pro, IFEval, IFBench, Multi-IF, and BFCLv4 form the core. GSM8K serves the two smaller models. AIME25 serves the two larger ones. The BF16 GGUF serves as the in-format ceiling. Results are the mean across five repeats.

Retention hits 97.1 percent, 96.5 percent, 97.4 percent, and 96.6 percent of the respective BF16 baselines. Across all four models QAD substantially improves the Q4_0 checkpoint.

Speed on actual devices

Decode throughput was measured on four targets. MacBook Pro and NucBox EVO-X2 use GPU inference. Samsung Galaxy S26 Ultra and Raspberry Pi 5 use Arm CPU inference. BF16 and F16 are shown as full-precision references where profiled.

The 230M and 350M QAD Q4_0 checkpoints match Q5_K_M quality within evaluation variance at a 4 to 33 percent higher decode throughput. The 1.2B and 2.6B QAD Q4_0 checkpoints match Q4_K_M quality at a 3 to 14 percent higher throughput. These checkpoints also match Unsloth's UD-Q4_K_XL where applicable for the 230M and 1.2B models.

Where to get them

The QAD GGUFs are available on Hugging Face today. Use the files with llama.cpp or any runtime that supports GGUF Q4_0 artifacts. A sample command loads the 350M model and prompts it with a question about C. elegans.

What the announcement does not cover

Training compute and data details are not disclosed. Comparison to Q3 or Q2 quantization levels is absent. Battery impact on the phone and thermal behavior on the Pi are not reported. Long-context performance at these sizes is not mentioned. The license terms for commercial deployment are not stated in the post.