LFM2.5-VL-DSpark

Overview

Liquid AI released LFM2.5-VL-DSpark on September 24, 2026, an experimental draft model for its LFM2.5-VL-3B vision-language model implementing speculative decoding: a small draft model proposes candidate tokens that the larger target model verifies. The ~280M-parameter draft uses a 4-layer attention-only decoder with a block size of 8-9, delivering exact speculative decoding (output identical to running the 3B target alone) with decode speedups up to 3.13x on-device and 2.66x on an H100, and end-to-end speedups of 2.62x on-device and 2.27x on GPU. It ships with day-one support for llama.cpp, MLX-VLM, and SGLang, and was evaluated on six vision tasks (VQA, text VQA, captioning, chart QA, reasoning, multi-turn conversation) under the MMSpec benchmark standard.

This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.

Key features

• ~280M-parameter draft model (4-layer attention-only decoder) pairing with LFM2.5-VL-3B
• Exact speculative decoding: output matches running the 3B target model alone
• Decode speedups up to 3.13x on-device, 2.66x on H100
• End-to-end speedups of 2.62x on-device, 2.27x on GPU
• Day-one support for llama.cpp, MLX-VLM, and SGLang
• Open weights in Safetensors and GGUF formats, unrestricted for download, fine-tuning, and deployment

Use cases

• Faster on-device vision-language inference without changing model outputs
• Lower-latency VLM deployment on llama.cpp, MLX-VLM, or SGLang
• Cost reduction for GPU-served vision-language inference via speculative decoding

Pricing

Free; open weights, unrestricted for download, fine-tuning, and deployment.

Pricing reflects the source record and may change. Confirm current terms with the provider.

Dates and sources

Released

24 Sept 2026

Record updated

25 Sept 2026

Record last checked

25 Sept 2026

• huggingface.co/blog/liquidai/lfm2-5-vl-dspark