Liquid AI releases LFM2.5-VL-DSpark to accelerate vision-language inference.

Liquid AI has released an experimental DSpark draft model engineered to accelerate inference for its LFM2.5-VL-3B vision-language model. Introduced in September 2026, the add-on implements a speculative decoding pipeline that preserves output fidelity while delivering substantial throughput gains across consumer hardware and datacenter GPUs. The DSpark architecture modifies speculative decoding for multimodal workloads by projecting visual patches and textual tokens into a unified representation prior to processing. A four-layer, attention-only drafter captures hidden states at fixed network depths and conditions on them to generate blocks of candidate tokens. This approach allows the inference engine to follow the same computational path as standard text models while maintaining identical dimensionality for cross-modal inputs. The draft model comprises approximately 280 million parameters, adding an 8.9 percent memory overhead to the base deployment. It was optimized through ten training epochs on a curated mixture of vision-language supervised fine-tuning data, with ablation studies confirming that four layers and a block size of nine maximize acceptance rates without diminishing returns. Benchmarks across diverse vision tasks demonstrate significant decoding acceleration. Deployed on Apple M5 Max silicon via MLX, the system yields decoding speedups between 2.30x and 3.13x, with end-to-end latency reductions ranging from 1.56x to 2.62x. Comparable results appear on M3 Ultra hardware using llama.cpp, where decoding improves by 1.57x to 2.14x. NVIDIA H100 configurations deliver more pronounced gains, with decoding throughput increasing from 2.66x up to 20.4x and end-to-end latency dropping by 1.64x to 2.27x. Liquid AI emphasizes that speculative decoding exclusively accelerates the autoregressive generation phase. Because vision-language models require heavy compute for vision encoding and prompt prefilling, particularly on edge devices, overall wall-time improvements are governed by Amdahl’s law. When prefill and encoding dominate latency, the raw decoding acceleration translates to proportionally modest system-wide speedups. The implementation supports day-one integration with SGLang, llama.cpp, and MLX-VLM. Developers can configure speculative parameters, including block sizes and attention backends, through standard command-line flags or Python APIs, with optimal block sizes dynamically aligned to hardware capabilities. The complete model family, available in both Safetensors and GGUF formats, is distributed via Hugging Face. Liquid AI positions the DSpark release as a foundational step toward efficient, production-ready deployment of multimodal AI across distributed and on-device environments.