Liquid AI LFM2.5-VL-3B Brings Faster Vision AI to Edge Devices
Liquid AI launches LFM2.5-VL-3B, a 3.1B edge vision model using about 3 GB, with faster screen reading, grounding, tool use and 11K-token GPU throughput.
Summary
On August 12, 2026, Liquid AI released LFM2.5-VL-3B on Hugging Face, calling the 3.1B vision-language model its most capable option for local hardware. It combines a SigLIP2 400M NaFlex vision encoder with the LFM2.5-2.6B backbone, pretrained on about 34T tokens and 4x more vision data from curated and synthetic caption, OCR, grounding and instruction sets. Liquid AI expanded the tokenizer to a 128K vocabulary for non-Latin scripts, then applied supervised fine-tuning with teacher distillation and Antidoom training, followed by multi-reward reinforcement learning. The direct-answer model targets low-latency applications and improves screen and UI understanding, natural-language object grounding, multi-image reasoning and function calling in text and vision-text tasks.
Liquid AI's normalized 0 to 100 tests used vLLM 0.26.0, recommended generation settings and non-reasoning mode. LFM2.5-VL-3B averaged 69.4, up from LFM2-VL-3B's 57.2, versus 52.0 and 59.7 for gemma-4-E2B-it and E4B-it, 64.6 and 69.4 for InternVL 3.5 2B and 4B, and 63.7 and 70.1 for Qwen3.5-2B and 4B. ScreenSpot-v2 desktop, mobile and web reached 78.7, 81.2 and 82.2, while RefCOCO averaged 87.9. IFEval scored 82.3, ToolSandbox 59.5 and BFCL V4 32.5.
The model occupies about 3 GB and decodes 228 tokens per second on an M5 Max, 116 on a Ryzen AI Max+ 395 and 20 on a Galaxy S26 Ultra. GPU throughput reaches about 11K tokens per second at high concurrency, nearly 1B daily output tokens on one H100. Day-one support includes llama.cpp, MLX, vLLM, SGLang and ONNX; downloads, a WebGPU browser demo, documentation and fine-tuning tutorials are available.
Positives
- ScreenSpot-v2 desktop, mobile and web scores rose to 78.7, 81.2 and 82.2 from 6.0, 7.6 and 2.5.
- RefCOCO grounding increased to 87.9 from 57.1, while BLINK rose to 61.5 from 50.2 and MuirBench to 58.3 from 34.9.
- ToolSandbox improved to 59.5 from 26.4, and BFCL V4 climbed to 32.5 from 20.5.
- About 11K output tokens per second at high concurrency is roughly twice the tested 4B-class models and approaches 1B daily tokens per H100.
- Roughly 3 GB of memory enables 228 tokens per second on M5 Max and 20 tokens per second on Galaxy S26 Ultra.
- 34T pretraining tokens, 4x more vision data and a 128K vocabulary broaden visual coverage and support non-Latin scripts.
Risks & concerns
- 69.4 overall ties InternVL 3.5 4B and trails Qwen3.5-4B's 70.1 aggregate score.
- CountBenchQA declined to 87.3 from LFM2-VL-3B's 92.2, while POPE slipped to 88.7 from 89.2.
- BFCL V4's 32.5 remains below gemma-4-E4B-it's 40.0 and Qwen3.5-4B's 53.6.
- Liquid AI conducted the reported benchmarks itself using non-reasoning mode, and the article provides no independent evaluation.
- Direct-answer operation prioritizes response speed and does not provide reasoning traces.