LFM2.5 Q4_0 release notes / what's new

Liquid AI has released updated 4-bit checkpoints (Q4_0 GGUFs) for the LFM2.5 model family, including the 230M, 350M, 1.2B-Instruct, and 2.6B parameter versions. These checkpoints utilize Quantization-Aware Distillation (QAD) to provide the memory and speed benefits of 4-bit quantization without the typical degradation in model quality.

Quantization-Aware Distillation (QAD) Performance

Quantization-Aware Distillation (QAD) allows a high-precision teacher model to be distilled into a quantized student model. This process recovers 97% of the average BF16 accuracy that is typically lost during standard quantization.

Benchmark results across reasoning, instruction-following, tool use, and agentic capabilities—including GPQA Diamond, MMLU-Pro, IFEval, IFBench, Multi-IF, and BFCLv4—demonstrate that QAD checkpoints retain a high percentage of their BF16 baseline performance:

  • LFM2.5-230M: 97.1% retention
  • LFM2.5-350M: 96.5% retention
  • LFM2.5-1.2B-Instruct: 97.4% retention
  • LFM2.5-2.6B: 96.6% retention

For mathematical evaluations, the 230M and 350M models were tested on GSM8K, while the 1.2B-Instruct and 2.6B models were tested on AIME25.

Edge Hardware Throughput and Efficiency

QAD Q4_0 checkpoints provide significant throughput improvements over higher-precision quantized versions while maintaining comparable quality. Testing was conducted across MacBook Pro and NucBox EVO-X2 (GPU inference) and Samsung Galaxy S26 Ultra and Raspberry Pi 5 (Arm CPU inference).

Key performance findings include:

  • LFM2.5-230M and 350M: These models match the quality of Q5_K_M checkpoints but offer 4-33% higher decode throughput.
  • LFM2.5-1.2B and 2.6B: These models match the quality of Q4_K_M checkpoints and provide 3-14% higher throughput.
  • External Comparison: QAD Q4_0 checkpoints match the performance of Unsloth's UD-Q4_K_XL (for the 230M and 1.2B models), which is a high-performance post-training quantization checkpoint.

Deployment and Usage

The QAD GGUFs are compatible with llama.cpp and any runtime that supports GGUF Q4_0 artifacts.

Implementation Example

Developers can run the LFM2.5-350M model using the following command:

llama-cli -hf LiquidAI/LFM2.5-350M \
  --hf-file LFM2.5-350M-QAD-Q4_0.gguf \
  -p "What is C. elegans?"

Available Models

Checkpoints are available on Hugging Face for the following models:

  • LFM2.5-230M
  • LFM2.5-350M
  • LFM2.5-1.2B-Instruct
  • LFM2.5-2.6B

Sources

Related

  • Dispatch
    LFM2.5-DSpark Release Notes / What's New

    Liquid AI has released DSpark draft model checkpoints for the LFM2.5 family, enabling up to 3.18x faster inference on GPUs and 2.87x on-device without altering output quality.

  • Dispatch
    LFM2.5-2.6B release notes / what's new

    Liquid AI has released LFM2.5-2.6B, a small, high-performance model designed for on-device agents with best-in-class tool use and instruction following capabilities.

  • Dispatch
    Quantization-Aware Healing enables a 4-bit LLM that outperforms its full‑precision original

    Hugging Face introduced Quantization‑Aware Healing (QAH), a method that compresses a GPT‑OSS 120B model to 60B parameters and 4‑bit precision while achieving higher accuracy than the original full‑precision checkpoint on most benchmarks.

  • Dispatch
    LFM2.5-VL-3B release notes / what's new

    Liquid AI has released LFM2.5-VL-3B, a 3.1B parameter vision-language model optimized for edge devices, featuring improved screen understanding, grounding, and tool calling.

  • Dispatch
    Unsloth Dynamic 3.0 GGUF Release

    Unsloth has released Dynamic 3.0 GGUFs, featuring improved quantization methodology that delivers over 10% better top-1% accuracy for Qwen3.8-27B compared to other providers at the same model size.