Source
LFM2.5 Q4\_0 Checkpoints from Quantization-Aware Distillation
Liquid AI released updated 4-bit QAD Q4_0 checkpoints for four LFM2.5 models. These use Quantization-Aware Distillation to maintain high accuracy while reducing memory usage and increasing speed.
Why it matters
This allows developers to run efficient, high-quality AI models on edge devices like smartphones and Raspberry Pis without significant performance loss.
The details
- QAD recovers 97% of average accuracy lost through quantization.
- Checkpoints are compatible with llama.cpp and other GGUF Q4_0 runtimes.
- Tested on hardware including Samsung Galaxy S26 Ultra and Raspberry Pi 5.
Show entities and relationshipsHide entities and relationships
In this article
Products
Technologies
Key connections
Liquid AI owns LFM2.5-230M
Liquid AI developed and released LFM2.5-230M with QAD Q4_0 checkpoints.
Liquid AI owns LFM2.5-350M
Liquid AI developed and released LFM2.5-350M with QAD Q4_0 checkpoints.
Liquid AI is a partner of Hugging Face
Liquid AI hosts and releases its QAD GGUF checkpoints on Hugging Face.
Unsloth owns UD-Q4_K_XL
Unsloth developed the UD-Q4_K_XL post-training quantization checkpoint.
Samsung owns Samsung Galaxy S26 Ultra
Samsung manufactures the Galaxy S26 Ultra smartphone.
Raspberry Pi owns Raspberry Pi 5
Raspberry Pi manufactures the Raspberry Pi 5 single-board computer.
Show 20 more connectionsShow fewer connections
LFM2.5-230M is built with Quantization-Aware Distillation
LFM2.5-230M Q4_0 checkpoints were trained using Quantization-Aware Distillation.
LFM2.5-350M is built with Quantization-Aware Distillation
LFM2.5-350M Q4_0 checkpoints were trained using Quantization-Aware Distillation.
LFM2.5-1.2B-Instruct is built with Quantization-Aware Distillation
LFM2.5-1.2B-Instruct Q4_0 checkpoints were trained using Quantization-Aware Distillation.
LFM2.5-2.6B is built with Quantization-Aware Distillation
LFM2.5-2.6B Q4_0 checkpoints were trained using Quantization-Aware Distillation.
Liquid AI owns LFM2.5-1.2B-Instruct
Liquid AI developed the LFM2.5-1.2B-Instruct language model.
UD-Q4_K_XL is built with Post-Training Quantization
UD-Q4_K_XL is an external post-training quantization checkpoint.
LFM2.5-230M competes with UD-Q4_K_XL
LFM2.5-230M QAD Q4_0 matches the quality of Unsloth's UD-Q4_K_XL.
LFM2.5-1.2B-Instruct competes with UD-Q4_K_XL
LFM2.5-1.2B-Instruct QAD Q4_0 matches the quality of Unsloth's UD-Q4_K_XL.
Quantization-Aware Distillation competes with Post-Training Quantization
Quantization-Aware Distillation is compared directly against post-training quantization methods.
LFM2.5-350M uses llama.cpp
LFM2.5-350M GGUF checkpoints can be executed using llama.cpp runtime.
llama.cpp supports GGUF Q4_0 artifact execution.
MacBook Pro uses GPU Inference
MacBook Pro was evaluated using GPU inference.
NucBox EVO-X2 uses GPU Inference
NucBox EVO-X2 was evaluated using GPU inference.
Samsung Galaxy S26 Ultra uses Arm CPU Inference
Samsung Galaxy S26 Ultra was evaluated using Arm CPU inference.
Raspberry Pi 5 uses Arm CPU Inference
Raspberry Pi 5 was evaluated using Arm CPU inference.
LFM2.5-230M uses MacBook Pro
LFM2.5-230M decode throughput was profiled on MacBook Pro.
LFM2.5-230M uses NucBox EVO-X2
LFM2.5-230M decode throughput was profiled on NucBox EVO-X2.
LFM2.5-230M uses Samsung Galaxy S26 Ultra
LFM2.5-230M decode throughput was profiled on Samsung Galaxy S26 Ultra.
LFM2.5-230M uses Raspberry Pi 5
LFM2.5-230M decode throughput was profiled on Raspberry Pi 5.
Liquid AI owns LFM2.5-2.6B
Liquid AI developed the LFM2.5-2.6B language model.
Related events
Release of QAD Q4_0 GGUF Checkpoints for LFM2.5 Models
Get the weekly recap
The stories like this one, picked and explained — once a week, straight to your inbox.