LFM2.5-VL-DSpark explained.

DIRECT ANSWER

What is LFM2.5-VL-DSpark?

LFM2.5-VL-3B-DSpark is an experimental draft model published by Liquid AI on September 24, 2026. It is a companion to LFM2.5-VL-3B, not a new general-purpose image model. The target processes visual and text inputs, while a lightweight drafter predicts several candidate output tokens. The target then checks those guesses in fewer verification passes. Source: Liquid AI release.

Think of the parent model as the image reader and final authority; DSpark is the draft assistant for the output stage. This distinction matters for search queries such as “LFM2.5-VL-DSpark image understanding”: the parent's image understanding determines answer quality. DSpark primarily changes inference throughput, not the vision model's abilities.

CHECKPOINT FACTS

LFM2.5-VL-3B-DSpark specifications

Data provenance: official drafter model card. Parameters are not a VRAM requirement and do not include the memory used by the target, image encoder, KV cache, runtime and loading overhead.

ARCHITECTURE

How does speculative decoding work for vision-language models?

By the time tokens reach the tapped layers, text and visual patches share the model's hidden-state representation. This is why a similar block-drafting technique can work for both text and multimodal output. In greedy decoding, exact verification preserves the target output; with matched stochastic sampling, the target output distribution is preserved. The system can still produce wrong descriptions if the target does—this is not an accuracy upgrade. Source: published DSpark methodology.

MEASURED BY LIQUID AI / NOT SXF

LFM2.5-VL-DSpark benchmarks: how fast is it?

Liquid AI benchmarked the same LFM2.5-VL-3B target with and without DSpark. Figures below are the highest task-specific values published for the named hardware; they are not averages and should not be compared with unrelated models' image-quality scores.

The six evaluation task families include general visual question answering (GQA), image captioning (COCO), text VQA, chart understanding (CharXiv), complex reasoning (MMMU-Pro) and multi-turn conversation. The checkpoint card publishes a task-by-task table. For example, on an M5 Max, COCO reached 3.13× decode speedup, while MMMU-Pro reached the top 2.62× end-to-end speedup; these are different tasks. Inspect the full official benchmark table ↗.

SXF EXPLANATORY TOOL

Why does 3× decoding not mean 3× faster vision inference?

Image encoding and prompt prefill still take time. Explore this illustrative Amdahl's-law model to see how the amount of baseline time spent decoding limits end-to-end benefit. This tool is not a measured DSpark benchmark and makes no prediction of memory use or real deployment speed.

SXF-derived arithmetic only. Ignore additional drafter overhead for this simplified upper-bound illustration. Unlike the hardware results above, this slider is not a vendor measurement.

LOCAL DEPLOYMENT REALITY

How much RAM or VRAM does LFM2.5-VL-DSpark need?

There is no universal minimum GPU VRAM or Mac unified-memory figure in the sources verified for this page. The drafter is only one part of the deployment. You must also load the LFM2.5-VL-3B target, its vision projector or encoder, KV cache, images, and runtime buffers. The stated 8.9% increase refers to parameter count, not guaranteed total runtime memory.

What to check before local deployment

1. Use a matching target and drafter checkpoint; the GGUF drafter shares token embeddings and the LM head with the loaded target.
2. Calculate the combined weight footprint plus vision encoder, runtime, KV cache, context length, resolution and batch overhead.
3. Choose a compatible runtime and use its own loading/VRAM measurements. Do not assume the 16-bit benchmark gains transfer to a 4-bit build.
4. Record peak memory and latency with and without drafting on the same images and prompt lengths.

PRACTICAL SETUP / OFFICIAL COMMANDS

How to run LFM2.5-VL-DSpark locally

Choose the runtime that matches your hardware. These examples are adapted from the official Liquid AI model cards and require appropriate runtime installation, model downloads, compatible dependencies and sufficient memory. The drafter alone cannot answer an image prompt.

Option 1: Apple silicon — MLX-VLM

Liquid AI lists MLX-VLM v0.7.2 or newer and greedy decoding for this DSpark integration.

python -m mlx_vlm.generate \
  --model LiquidAI/LFM2.5-VL-3B \
  --draft-model LiquidAI/LFM2.5-VL-3B-DSpark \
  --draft-block-size 8 \
  --image /path/to/image.jpg \
  --prompt "Describe this image." \
  --max-tokens 256 \
  --temperature 0

Option 2: NVIDIA GPU — SGLang

The current model card specifies SGLang v0.5.19 or newer with DSpark integration. Check the installed build and supported attention backend first.

python -m sglang.launch_server \
  --model-path LiquidAI/LFM2.5-VL-3B \
  --speculative-algorithm DSPARK \
  --speculative-draft-model-path LiquidAI/LFM2.5-VL-3B-DSpark \
  --speculative-draft-attention-backend flashinfer \
  --speculative-dspark-block-size 9 \
  --disable-radix-cache --mem-fraction-static 0.8 --port 30000

Connect a compatible client to the local OpenAI-style endpoint at http://localhost:30000/v1. Compare it to the same target with the speculative flags removed. Do not expose a development server on a public network without authentication and deployment safeguards.

Option 3: llama.cpp — matching GGUF pair

Liquid AI provides a separate F16 drafter GGUF, used with the corresponding target GGUF and vision files.

llama-server -hf LiquidAI/LFM2.5-VL-3B-GGUF:F16 \
  -hfd LiquidAI/LFM2.5-VL-3B-DSpark-GGUF:F16 \
  --spec-type draft-dspark \
  --spec-draft-n-max 8 --spec-draft-n-min 0 \
  -ngl 99 -ngld 99 -fa on

Keep each drafter paired with the correct target format. Runtime flags and model availability can change; the maintained official model card takes precedence over commands reproduced here.

REAL WORKFLOWS

Vision-language use cases: where DSpark helps—and where it does not

DSpark's potential is highest when the existing LFM2.5-VL-3B model is already accurate enough for the task and output generation occupies a meaningful share of total latency. The following are proposed workflows to evaluate, not SXF-tested performance results.

DSpark versus another small vision model

Liquid AI's parent-model evaluation publishes comparative quality figures. For example, it reports 80.7 average ScreenSpot-v2 for LFM2.5-VL-3B and 61.5 BLINK for multi-image reasoning in its evaluation setup. Those are target-model results, not DSpark accuracy scores or independent SXF measurements.

HONEST LIMITATIONS

Limitations, accuracy and licensing

• Not a standalone VLM: requires a compatible LFM2.5-VL-3B target plus the vision stack.
• Only decode is accelerated: vision encoding and prefill still dominate some edge workloads. Gains vary with prompt/image complexity and output length.
• Acceptance rate matters: at higher sampling temperatures the target and drafter can diverge more often, potentially reducing throughput.
• No extrapolation to quantized models: released benchmark results use 16-bit inference and do not establish 4-bit or other quantized acceleration.
• Vendor benchmarks: reported gains are from Liquid AI; SXF has not independently reproduced them.
• License is not a legal conclusion: Hugging Face labels the checkpoint lfm1.0. Read the complete license terms for your use case.

SEARCH QUESTIONS

LFM2.5-VL-DSpark FAQs

EVIDENCE REGISTER

Official sources and verification notes

Last editorial source review: November 1, 2026. The release date is September 24, 2026. Every performance or size claim above is grounded in official model documentation; the Amdahl calculator and deployment checklist are SXF's own analysis.

Explore related SXF resources: AI models · Open-source AI model guide · Run AI models locally · Quantization guide · Original SXF release signal.