Liquid AI Releases LFM2.5-VL-450M: a 450M-Parameter Vision-Language Model with Bounding Box Prediction, Multilingual Support, and Sub-250ms Edge Inference

Read the full articleLiquid AI Releases LFM2.5-VL-450M: a 450M-Parameter Vision-Language Model with Bounding Box Prediction, Multilingual Support, and Sub-250ms Edge Inference on MarkTechPost

↗

What Happened

Liquid AI just released LFM2.5-VL-450M, an updated version of its earlier LFM2-VL-450M vision-language model. The new release introduces bounding box prediction, improved instruction following, expanded multilingual understanding, and function calling support — all within a 450M-parameter footprint

Our Take

450 million parameters for a Vision-Language Model is respectable, but the real win here is the efficiency. achieving sub-250ms edge inference is what matters for deployment, not just the parameter count. that's where the real engineering challenge lies: squeezing high-level vision and language understanding onto constrained hardware.

adding bounding box prediction and multilingual support is expected, but packaging it all into a single, efficient model is the difficult part. if they can maintain that latency while supporting complex instructions, it’s a major step toward genuinely deployable multimodal AI, especially for edge devices.

What To Do

Benchmark LFM2.5-VL-450M inference speed on target edge hardware (e.g., a standard mobile GPU).

Builder's Brief

Who

mobile and edge ML engineers building vision pipelines on constrained hardware

What changes

new lightweight VLM option with bounding box prediction that could replace heavier detection stacks on-device

When

now

Watch for

whether LFM2.5-VL-450M gets wrapped into ONNX or CoreML conversion tooling by the community

What Skeptics Say

Sub-250ms inference claims depend on specific hardware configurations that most edge deployments don't match; accuracy on bounding box prediction at 450M parameters will degrade significantly on out-of-distribution objects. Liquid AI's architecture novelty hasn't translated into broad ecosystem adoption yet.

Cited By

MarkTechPost Liquid AI Releases LFM2.5-VL-450M: a 450M-Parameter Vision-Language Model with Bounding Box Prediction, Multilingual Support, and Sub-250ms Edge Inference