# FastVLM vs LFM2.5-VL

FastVLM (0.5B–7B) versus Liquid AI’s LFM2.5-VL (450M and 3B, 2026): grounding, tool calling, multi-image, browser and llama.cpp support, reported speed and licenses.

Canonical URL: https://aiidelist.com/blog/fastvlm-vs-lfm2-5-vl

Language: en

Published: 2026-09-24

Updated: 2026-09-24

FASTVLM VS LFM2.5-VL

LFM2.5-VL is Liquid AI’s 2026 edge vision family: a 450M model released in April and a 3B model in August that adds grounding, screen understanding and function calling. FastVLM is Apple’s 2025 family built around encoder efficiency. Here is what each documents.

[All comparisons](/fastvlm#comparisons)

| Feature | FastVLM | LFM2.5-VL |
| --- | --- | --- |
| Developer | Apple researchers | Liquid AI |
| Sizes and release | 0.5B · 1.5B · 7B (2025) | 450M (8 Apr 2026) · 3B = 2.6B LM + SigLIP2 400M NaFlex (12 Aug 2026) |
| Object grounding (boxes) | Not a documented feature | Yes — RefCOCO-avg 87.9 (3B), 81.28 RefCOCO-M (450M), Liquid-reported |
| Screen / UI understanding | Strong document/OCR scores; no ScreenSpot figures published | ScreenSpot-v2 78.7 desktop · 81.2 mobile · 82.2 web (3B, Liquid-reported) |
| Function / tool calling | Not a documented feature | Yes |
| Multi-image | Single image in official examples | Yes |
| Browser (WebGPU) | 0.5B ONNX (this site) | LFM2.5-VL-3B-ONNX and an official WebGPU Space |
| Runtimes | Transformers · MLX (official app) · Core ML · Transformers.js | llama.cpp · MLX · vLLM · SGLang · ONNX · LEAP (day-one, Liquid-stated) |
| Reported on-device speed | Up to 85× faster TTFT vs LLaVA-OneVision-0.5B at 1152 px; no tokens/s published | 3B: 228 tok/s on M5 Max, 20 tok/s on Galaxy S26 Ultra, ≈3 GB memory (Liquid) |
| License | Apple ML Research license | LFM Open License (check the model card for the current version) |

Checked 2026-09-17

## Decision summary

- LFM2.5-VL documents more capabilities—boxes, tool calls, multi-image, GGUF—so it is the easier fit for agent-style edge apps.
- FastVLM remains the reference for encoder efficiency at high resolution and has the tightest Apple-platform integration (MLX app, Core ML assets).
- Both have browser builds; the 3B LFM model needs roughly three times the memory of FastVLM-0.5B.

## Choose FastVLM when

- You want the smallest high-resolution model for a browser or iPhone demo
- Your pipeline is MLX or Core ML on Apple hardware
- Latency to first token on large images is the metric you optimise

## Choose LFM2.5-VL when

- You need bounding boxes, UI grounding or function calling
- You ship with llama.cpp/GGUF across Android, Windows and Linux
- You want Liquid’s published throughput numbers as a starting point

## How to read this table

Every number above is vendor-reported on the vendor’s own hardware and evaluation harness. FastVLM’s 85× figure is a TTFT ratio at one resolution; Liquid’s tokens/s are decode rates on specific devices. Neither is a general ranking.

## Questions

**Does FastVLM support bounding boxes?**

Apple’s model cards do not document grounding outputs. You can ask it to describe positions in text, but there is no evaluated box format as in LFM2.5-VL.

**Which runs on Android?**

LFM2.5-VL ships GGUF for llama.cpp and Liquid publishes phone numbers. FastVLM has no official Android path; community GGUF or ONNX ports may exist but are not covered here.

## Sources

- [LFM2.5-VL-3B announcement](https://huggingface.co/blog/LiquidAI/lfm2-5-vl-3b)
- [LFM2.5-VL-450M announcement](https://www.liquid.ai/blog/lfm2-5-vl-450m)
- [FastVLM paper](https://arxiv.org/abs/2412.13303)

## Continue

- [On-device VLM hub](/blog/on-device-vlm-comparison)
- [FastVLM vs SmolVLM2](/blog/fastvlm-vs-smolvlm2)
- [FastVLM models](/blog/fastvlm-models)