What is LFM2.5-DSpark

LFM2.5-DSpark is a model optimization technology developed by Liquid AI that achieves high-speed inference, based on the traditional LFM2.5 architecture. This technology achieves 3.2 times faster speed and enhances on-device execution. It is particularly optimized for use cases such as data extraction, tool usage, and structured output (source: https://huggingface.co/blog/LiquidAI/lfm25-dspark).

Technical Details

LFM2.5-DSpark supports the following model sizes and formats:

  • 350M parameters: LFM2.5-350M and LFM2.5-350M-GGUF
  • 1.2B parameters: LFM2.5-1.2B-Instruct and LFM2.5-1.2B-Instruct-GGUF
  • 2.6B parameters: LFM2.5-2.6B and LFM2.5-2.6B-MLX
  • 8B parameters: LFM2.5-8B-A1B

Each model enables cross-platform deployment through Quantized format (GGUF, ONNX, MLX) and ONNX Runtime (source: not mentioned in official documentation).

Architecture and Optimization

LFM2.5-DSpark improves the LFM2 architecture based on extended pre-training and reinforcement learning, achieving efficient inference. In particular, optimization for Apple Silicon using the MLX framework enables high-speed execution on Mac devices. It also includes frameworks that support CPU offloading and hardware acceleration (vLLM, llama.cpp) (source: not mentioned in official documentation).

Inference Framework

LFM2.5-DSpark works with the following frameworks:

  • Transformers: allows direct access to the model internals
  • vLLM: high-throughput production deployment using GPU
  • llama.cpp: cross-platform inference that supports CPU offloading
  • MLX: optimized for Apple Silicon
  • LM Studio: desktop application for local LLM execution

In particular, it has achieved 220 tokens/second decoding speed on M5 Max and 113 tokens/second on Ryzen AI Max+ 395 (source: https://huggingface.co/LiquidAI/LFM2.5-2.6B).

What is Possible

By utilizing LFM2.5-DSpark, the following implementations are possible:

  • Data extraction and structured output: optimized for tool usage and RAG (Retrieval-Augmented Generation)
  • Vision-language tasks: enables image caption generation and object detection in collaboration with LFM2.5-VL-450M
  • On-device inference: enables execution in low-resource environments using MLX and ONNX
  • Multi-platform deployment: flexible deployment on CPU, GPU, and Apple Silicon (source: not mentioned in official documentation).

Summary

  • LFM2.5-DSpark’s acceleration technology achieves 3.2 times faster inference speed
  • MLX and ONNX enable deployment on Apple Silicon and cross-platform
  • Model configuration optimized for data extraction and tool usage use cases
  • vLLM and llama.cpp enable high-efficiency inference on GPU and CPU
  • Collaboration with vision-language models (LFM2.5-VL-450M) enables multi-modal processing.