Liquid AI builds small models designed for specialization and on-device deployment. LFM3 brings a new tri-hybrid architecture and a shared backbone for text, vision, audio, omni, and decision models.

The model featured here, LFM3-VL-2B, combines the LFM3 language backbone with an in-house vision encoder for text, image, and video input. It is designed for reading documents, charts, and app screens, answering questions about video, finding objects, and reasoning about their spatial relationships.

Model overview

• 1.92B parameters: a 1.73B language model, 164M vision encoder, and 25M projector.
• 128K context: input and output share the context window. This is the vision-language model's context limit, not the text model's 256K context.
• Tri-hybrid language backbone: 24 layers combining gated short convolutions, sliding-window attention, and global attention.
• Visual understanding: supports images and video; the upstream card describes video sampling at 2 fps. Video availability in local applications depends on runtime support.
• Reasoning and tool use: reasons before answering and supports structured tool calls. The local artifact exposes thinking and thinking-history preservation controls.

Liquid AI-reported benchmark results

Selected category averages from Liquid AI's provisional evaluation across 62 vision benchmarks:

Category	LFM3-VL-2B	LFM2.5-VL-3B
Document, OCR & chart	83.9	71.2
Math & logical reasoning	68.8	40.6
Visual grounding	77.3	68.3
Embodied reasoning	64.4	48.0
GUI understanding	83.1	66.0
Video	61.2	41.3
Agentic	42.5	21.1
Average across all 62 benchmarks	71.4	57.6

Source: Liquid AI's supplied LFM3-VL-2B draft model card. Results are provisional. Models were evaluated with vLLM 0.27.1 using their recommended generation parameters when available. LFM3-VL-2B uses reasoning mode; LFM2.5-VL-3B has no reasoning mode. Category scores are unweighted means of their benchmarks; the overall average is the unweighted mean across all 62 benchmarks. Scores are normalized to 0–100. These are upstream results, not measurements of the downloadable GGUF quantization.

Local use and specialization

The linked LFM3-VL-2B artifact provides GGUF downloads for local inference. Liquid AI positions the LFM3 family for single-GPU fine-tuning and specialization to application data and deployment environments. Actual memory use, usable context length, and supported modalities depend on the checkpoint and runtime.

The local artifact uses temperature 0.1, top-k 50, and repetition penalty 1.1, with thinking enabled and preservation of prior thinking disabled by default.

License

The supplied model card specifies the LFM Open License v1.0.