---

library_name: llama.cpp

base_model: LiquidAI/LFM2.5-VL-3B-DSpark

license: other

license_name: lfm1.0

license_link: LICENSE

pipeline_tag: image-text-to-text

tags:

• speculative-decoding
• dspark
• lfm2
• lfm2-vl
• draft-model
• gguf

---

<div align="center">

<img

src="https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/2b08LKpev0DNEk6DlnWkY.png"

alt="Liquid AI"

style="width: 100%; max-width: 100%; height: auto; display: inline-block; margin-bottom: 0.5em; margin-top: 0.5em;"

/>

<div style="display: flex; justify-content: center; gap: 0.5em; margin-bottom: 1em;">

<a href="https://playground.liquid.ai/"><strong>Try LFM</strong></a> •

<a href="https://docs.liquid.ai/lfm/getting-started/welcome"><strong>Docs</strong></a> •

<a href="https://discord.com/invite/liquid-ai"><strong>Discord</strong></a>

</div>

</div>

LFM2.5-VL-3B-DSpark-GGUF

GGUF build of LiquidAI/LFM2.5-VL-3B-DSpark for llama.cpp.

This is a standalone draft sidecar: it carries only the drafter (4 attention layers, a Markov head, a confidence head, and block size 9.) Token embeddings and the LM head are shared from the target model at load time, so it must be paired with a LFM2.5-VL-3B-GGUF target file.

Find more information about LFM2.5-VL-DSpark in our blog post.

📦 Files

| file | size | notes |

| ------------------------------- | ------ | ---------------------------------------------------------------------------------------------------------------------- |

| LFM2.5-VL-3B-DSpark-F16.gguf | 567 MB | Standalone F16 drafter; pair with the target model and vision projector from LiquidAI/LFM2.5-VL-3B-GGUF. |

All inference numbers for this release use 16-bit processing for both the vision encoder and language backbone.

🏃 How to run (llama.cpp)

Run the target model with the DSpark drafter:

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

The drafter was trained with block size 9. For Apple silicon, we recommend block size 8 through --spec-draft-n-max 8.

Speculative decoding is exact under greedy decoding: the target verifies every proposed token, so the generated output equals the target model running alone. The llama.cpp timing logs report the draft acceptance rate.

Other LFM2.5-VL-3B-DSpark formats:

| Draft | Target |

| ------------------------------------------------------------------------------------ | ---------------------------------------------------------------------- |

| LFM2.5-VL-3B-DSpark | LFM2.5-VL-3B |

| LFM2.5-VL-3B-DSpark-GGUF | LFM2.5-VL-3B-GGUF |

📊 Acceptance and benchmarks

See LiquidAI/LFM2.5-VL-3B-DSpark for model details and acceptance-length and throughput benchmarks across six vision-language tasks on NVIDIA H100 and Apple silicon, including MLX-VLM, llama.cpp, and SGLang results.

📬 Contact

• Got questions or want to connect? Join our Discord community
• If you are interested in custom solutions with edge deployment, please contact our sales team.

Citation

@article{liquidAI2026VL3B,
  author  = {Liquid AI},
  title   = {LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge},
  journal = {Liquid AI Blog},
  year    = {2026},
  note    = {www.liquid.ai/blog/lfm2-5-vl-3b},
}

@article{liquidAI2026vldspark,
  author = {Liquid AI},
  title = {LFM2.5-VL-DSpark: Accelerating vision-language models on edge and beyond},
  journal = {Liquid AI Blog},
  year = {2026},
  note = {www.liquid.ai/blog/lfm2-5-vl-dspark},
}