Local VLM Pushes Performance Boundaries
LiquidAI has released LFM2.5-VL-3B, a 3.1 billion parameter vision-language model (VLM) that demonstrates impressive performance, notably outperforming Google's Gemma-4 E4B and achieving a dramatic leap in screen understanding capabilities. The model's architecture allows for full local execution, supporting frameworks like llama.cpp, MLX, vLLM, and even offering a WebGPU demo, making advanced AI accessible without relying on cloud infrastructure.
Performance benchmarks show LFM2.5-VL-3B achieving a score of 69.4 compared to Gemma-4 E4B's 59.7, and it also edges out the Qwen3.5-4B model. This competitive edge is particularly significant given the smaller parameter count of LFM2.5-VL-3B. The model's most striking improvement is in screen understanding, where it jumped from a baseline of 2.5 to an exceptional 82.2 on the ScreenSpot-v2 Web benchmark. This represents a more than thirty-fold increase, indicating a profound improvement in its ability to interpret and interact with visual interfaces.
Speed is another critical factor for local AI. LFM2.5-VL-3B processes data at 228 tokens per second on an Apple M5 Max chip, a respectable throughput for on-device processing. Even on mobile hardware, it achieves 20 tokens per second on a Galaxy S26 Ultra, suggesting practical applications for consumer devices. The model also integrates function calling and object grounding capabilities, enhancing its utility for real-world tasks.
The Business Angle: Local AI Redefines 'AI Feature'
Beyond raw technical performance, the release of LFM2.5-VL-3B highlights a significant shift in the business model for AI-powered features, particularly for document and screen-based tasks. Traditionally, integrating AI meant recurring API costs from cloud providers. This new generation of capable local models flips that paradigm. For businesses, implementing an 'AI feature' can now transition from an ongoing operational expense to a one-time engineering investment. This not only offers cost predictability but also addresses critical data privacy and security concerns, as sensitive information no longer needs to leave the local environment for processing.
This move towards local processing is not about replacing large, cloud-based models entirely. Those models will continue to lead in complex reasoning tasks requiring vast computational resources and extensive world knowledge. Instead, LFM2.5-VL-3B and similar models are making the 'small model' category genuinely viable and powerful for a wide array of practical applications. Think of it less like a replacement for a supercomputer and more like fitting a highly capable, specialist tool into your pocket. This democratizes AI capabilities, allowing smaller companies and individual developers to build sophisticated AI features without prohibitive ongoing costs or data governance hurdles.
Technical Advancements Driving Local VLM Success
The leap in screen understanding performance is a testament to advancements in VLM architectures and training methodologies. The ScreenSpot-v2 benchmark, designed to evaluate a model's ability to comprehend visual interfaces, requires nuanced understanding of elements like buttons, text fields, icons, and their spatial relationships. Achieving an 82.2 score suggests LFM2.5-VL-3B can accurately identify, classify, and potentially interact with these components with high fidelity. This capability is crucial for developing AI agents that can navigate and operate software, automate UI testing, or provide intelligent assistance within digital environments.
The integration of function calling and object grounding further elevates LFM2.5-VL-3B's practical utility. Function calling allows the model to trigger specific actions or retrieve information from external systems based on its understanding of visual input. Object grounding enables the model to precisely identify and locate specific objects within an image or screen. Together, these features pave the way for more interactive and agentic AI applications that can not only perceive but also act upon their environment. The ability to run these advanced functions locally means that applications requiring real-time interaction and high data sensitivity can now be built with greater autonomy and security.
The broad framework support—llama.cpp for CPU inference, MLX for Apple Silicon, vLLM for efficient serving, and WebGPU for browser-based applications—ensures that LFM2.5-VL-3B can be deployed across a wide spectrum of hardware. This versatility is key to unlocking the potential of local AI, allowing developers to choose the most appropriate inference engine based on their target platform and performance requirements. The inclusion of a WebGPU demo, in particular, lowers the barrier to entry for web developers looking to integrate VLM capabilities directly into their applications without complex backend infrastructure.