LFM2.5-VL is Liquid AI's family of compact vision-language models designed for efficient multimodal understanding and deployment on edge devices. The family combines Liquid's hybrid convolution-and-attention language architecture with SigLIP2-based vision encoders, supporting image understanding, instruction following, and lightweight agentic workflows with dynamic image-token allocation. The models are based on the LFM2 technical report and are designed to provide different capability and efficiency points for resource-constrained multimodal applications.

Technical reference: LFM2 Technical Report — arXiv:2511.23404

LFM2.5-VL-450M

LFM2.5-VL-450M is the smallest variant, combining the LFM2.5-350M language backbone with an 86M-parameter SigLIP2 NaFlex vision encoder. Its compact footprint, dynamic image-token allocation, multilingual support, and native processing of images up to 512×512 make it particularly suitable for highly constrained edge and embedded deployments.

Model Configuration:

• Reference implementation: transformers
• Original Weight: LFM2.5-VL-450M
• Vision Encoder: SigLIP2 NaFlex (86M)
• Language Model: LFM2.5-350M
• Resolution: 512×512
• Support Cooper version:
  • Cooper SDK: [2.5.4] + Shepherd_1.6.8
  • Cooper Foundry: [2.3] + Shepherd_1.6.8

Model	Device	Model Link
LFM2.5-VL-450M	CV72	Model_Link
LFM2.5-VL-450M	CV75	Model_Link

LFM2.5-VL-1.6B

LFM2.5-VL-1.6B increases language-model and visual-encoder capacity to provide a stronger balance between multimodal capability and inference efficiency. It is targeted at applications requiring more capable visual reasoning and instruction following while retaining the low-latency, edge-oriented design of the LFM2.5-VL family.

Model Configuration:

• Reference implementation: transformers
• Original Weight: LFM2.5-VL-1.6B
• Vision Encoder: SigLIP2 NaFlex (400M)
• Language Model: LFM2.5-1.2B-Base
• Resolution: 512×512
• Support Cooper version:
  • Cooper SDK: [2.5.4] + Shepherd_1.6.8
  • Cooper Foundry: [2.3] + Shepherd_1.6.8

Model	Device	Model Link
LFM2.5-VL-1.6B	N1-655	Model_Link
LFM2.5-VL-1.6B	X7	Model_Link
LFM2.5-VL-1.6B	CV7	Model_Link