LiquidAI Announces Release of LFM2.5-VL-3B Multimodal Model

LiquidAI has released LFM2.5-VL-3B, a multimodal variant of its LFM2.5 hybrid model family. Designed specifically for on-device deployment, the new model integrates text and image processing capabilities by combining the LFM2.5-2.6B language model backbone with a SigLIP2 NaFlex vision encoder.

The release marks an evolution of the previous LFM2-VL-3B, incorporating further mid- and post-training refinements. According to technical specifications provided by the developer, the model features a vocabulary size of 128,000 and a context length of 32,768 tokens. It supports a wide range of languages, including English, Chinese, Arabic, French, German, Japanese, and several others.

Enhanced Vision and OCR Capabilities

The LFM2.5-VL-3B architecture utilizes SigLIP2’s NaFlex technology for native resolution processing. For larger images, the system splits the input into non-overlapping 512×512 patches alongside a resized whole-image thumbnail.

The developers report improvements in two primary areas: grounding and Optical Character Recognition (OCR). The model demonstrates enhanced object detection when prompted with natural language queries and provides full-page OCR with layout annotation. LiquidAI suggests the model is optimized for high-throughput, low-latency tasks such as near-real-time object detection for automotive applications, batch processing of scanned documents, and on-device translation of signage and menus. However, the company notes the model is not intended for reasoning-intensive tasks, such as visual web design or technical blueprint analysis.

Performance Benchmarks and Hardware Compatibility

The model is engineered for efficiency, requiring approximately 3 GB of memory. In hardware-specific testing, LiquidAI reported decoding speeds of 228 tokens per second (tok/s) on an Apple M5 Max and 116 tok/s on an AMD Ryzen AI Max+ 395. The company also noted that the model can run on a Galaxy S26 Ultra, reaching speeds of 20 tok/s.

For high-concurrency enterprise environments, the model was tested on a single NVIDIA H100. In these settings, using vLLM, the model reportedly achieved a throughput of approximately 11,000 tokens per second, which equates to nearly 1 billion tokens per day.

Community Feedback and Unverified Observations

Following the release, early users on platforms such as Reddit have shared preliminary performance data. One user reported achieving 35 tokens per second on an A380 using Vulkan, while another noted the model utilized just over 3 GB of VRAM when running at a Q5km quantization with a 32k context. While these community reports suggest the model “just works” across various setups, these figures remain unverified by independent third-party benchmarks.