Original research brief
Liquid AI Releases QAD Q4_0 Checkpoints for LFM2.5 Models
Liquid AI released QAD Q4_0 checkpoints for four LFM2.5 models, claiming improved quality retention and throughput on edge hardware compared to standard…
OMNIASSIST / FIELD NOTESnews · source-led editorial
What this piece is grounded in
01
Liquid AI released QAD Q4_0 GGUFs for four LFM2.5 models.
02
QAD checkpoints retain 97.1%, 96.5%, 97.4%, and 96.6% of BF16 baseline performance.
03
The 230M and 350M QAD Q4_0 checkpoints match Q5_K_M quality within evaluation variance.
04
The 1.2B and 2.6B QAD Q4_0 checkpoints match Q4_K_M quality at higher throughput.
01 / FIELD NOTE
What Changed
Liquid AI has released QAD Q4_0 GGUFs, which are updated 4-bit checkpoints for four LFM2.5 models: 230M, 350M, 1.2B-Instruct, and 2.6B. These checkpoints were produced using quantization-aware distillation (QAD) rather than the standard post-training quantization (PTQ) approach. The company states that these QAD checkpoints allow developers to run the models at Q4_0 memory and speed without the usual quality drop. The files are available on Hugging Face and can be used with llama.cpp or any runtime that supports GGUF Q4_0 artifacts.
02 / FIELD NOTE
Why It Matters for Practical AI Work
The release addresses a common trade-off in edge AI: running smaller models with reduced memory footprint often comes with a loss in quality. Liquid AI reports that the QAD checkpoints retain 97.1%, 96.5%, 97.4%, and 96.6% of their respective BF16 baseline performance across the four models. Additionally, the 230M and 350M QAD Q4_0 checkpoints are said to match Q5_K_M quality within evaluation variance at a 4-33% higher decode throughput, while the 1.2B and 2.6B checkpoints match Q4_K_M quality at a 3-14% higher throughput. This could make it easier to deploy capable models on devices like phones and single-board computers.
03 / FIELD NOTE
What a Reader Can Verify
The QAD Q4_0 GGUFs are available on Hugging Face for the four named LFM2.5 models. The source article includes benchmark results comparing the QAD checkpoints against PTQ-produced GGUFs on a suite covering reasoning, instruction-following, tool use, and agentic capabilities. It also reports decode throughput measurements on MacBook Pro, NucBox EVO-X2, Samsung Galaxy S26 Ultra, and Raspberry Pi 5. Readers can download the files and test them with llama.cpp or other GGUF-compatible runtimes to confirm the reported quality and speed characteristics.
04 / FIELD NOTE
What Remains Unknown
The source does not specify the exact benchmark scores for each model, only the percentage of BF16 baseline performance retained. It also does not detail the hardware configurations used for the throughput measurements beyond the device names and whether GPU or CPU inference was used. The article does not mention pricing, licensing terms, or availability dates beyond the release. Additionally, the claim that QAD checkpoints match Q5_K_M or Q4_K_M quality is based on the company's own evaluation, and independent verification is not yet available.
Questions readers ask
What are QAD Q4_0 checkpoints?
QAD Q4_0 checkpoints are 4-bit quantized model files produced using quantization-aware distillation, a training method that aims to reduce quality loss compared to standard post-training quantization. Liquid AI released these for four LFM2.5 models, allowing them to run at lower memory and speed while retaining a higher percentage of the full-precision baseline performance.
How can developers use these checkpoints?
The QAD Q4_0 GGUFs are available on Hugging Face and can be used with llama.cpp or any runtime that supports GGUF Q4_0 artifacts. Developers can download the files for the specific LFM2.5 model they need and integrate them into their applications, potentially benefiting from improved quality and throughput on edge hardware.