Liquid AI Releases LFM2.5-VL-3B-DSpark, a Speculative Decoding Model for VLMs

Liquid AI has released 'LFM2.5-VL-3B-DSpark', a draft model for Speculative Decoding for its vision-language model (VLM) LFM2.5-VL-3B. This model is based on the DSpark approach by the research team from DeepSeek-AI and Peking University. With approximately 280M parameters, it significantly increases decoding speed while maintaining output quality at a parameter increase cost of 8.9% compared to the base model.

Performance and Hardware-Specific Improvements

• Apple M5 Max: Decoding speed improves by up to 3.13x when using MLX-VLM.
• NVIDIA H100: Decoding speed improves by up to 2.66x when using SGLang.
• Accepted Tokens: An average of 3.2 to 4.5 tokens are accepted per verification, supporting efficient inference.

Technical Architecture and Implementation

The DSpark approach uses a semi-autoregressive structure combining a parallel backbone and a sequential head (Markov Chain). Since image patches and text tokens are projected into the same dimension, the same inference algorithms as text models can be applied.

Supported Environments and Notes

• llama.cpp: Requires a build after PR #29339; block size 8 is recommended for Apple Silicon.
• SGLang: Requires v0.5.19 or higher, but support for LFM2-VL currently exists only in the main branch.
• MLX-VLM: Supported in v0.7.2 or higher, but currently only greedy sampling is available.
• License: Applies the LFM Open License v1.0, allowing commercial use only for companies with annual revenue under 10 million USD.