A new compact vision-language model from Liquid AI is challenging the assumption that bigger AI models always perform better. The LFM2.5-VL-3B, released on August 12, delivers competitive performance against models twice its size while consuming less computing power and running faster across phones, laptops, and cloud servers.

This development reflects a broader industry shift away from the "bigger is better" philosophy that has dominated artificial intelligence for the past five years. Instead of allocating extra processing power during inference like reasoning-focused models do, the LFM2.5-VL-3B answers questions directly without intermediate reasoning steps, keeping latency low for real-time applications. By avoiding the computational overhead of step-by-step thinking, the model achieves efficiency gains that make it practical for deployment on consumer devices.

How Does This Model Outperform Larger Competitors?

The LFM2.5-VL-3B contains 3.1 billion parameters, roughly one-quarter the size of some competing models. Yet it achieves remarkable results across multiple tasks. On screen understanding benchmarks, it scored 80.7 out of 100 on ScreenSpot-v2, far exceeding Google's Gemma-4-E4B model, which scored 51.2 despite being more than twice as large.

The model's improvements span four major capability areas. For function calling, a critical task where AI systems must decide which tools to use to solve problems, the model more than doubled its performance from 26.4 to 59.5 on the ToolSandbox benchmark. On grounding tasks, where the model must locate specific objects in images and draw bounding boxes around them, performance jumped 30 points from 57.1 to 87.9 on RefCOCO precision metrics. Multi-image reasoning also improved significantly, with BLINK scores climbing from 50.2 to 61.5.

What Makes This Model Practical for Real-World Use?

The efficiency gains matter because they determine whether AI can run on consumer devices or requires expensive cloud servers. The LFM2.5-VL-3B decodes 228 tokens per second on an Apple M5 Max processor and 116 tokens per second on an AMD Ryzen AI Max+ 395 processor, while using only about 3 gigabytes of memory. Most remarkably, it runs on a Galaxy S26 Ultra smartphone, decoding 20 tokens per second, making capable vision AI accessible without internet connectivity.

On graphics processing units (GPUs), the model returns its first token in roughly 34 milliseconds when processing a five-frame video clip, compared to around 200 milliseconds for larger Gemma models. At high concurrency levels, the model achieves approximately 11,000 output tokens per second on a single NVIDIA H100 GPU, roughly double the throughput of larger 4-billion-parameter models. This translates to nearly 1 billion output tokens per day on a single high-end GPU.

How to Deploy This Model Across Different Platforms

• Edge Devices: Run LFM2.5-VL-3B locally on phones and laptops using llama.cpp for CPU inference or MLX for Apple Silicon optimization, ensuring data privacy and eliminating cloud dependency.
• Production Servers: Deploy on data centers using vLLM or SGLang for GPU-accelerated serving, achieving high throughput for multiple simultaneous user requests.
• Cross-Platform Compatibility: Use ONNX format to run the model across diverse hardware accelerators, from specialized AI chips to standard processors, without rewriting code.
• Browser-Based Testing: Try the model directly in a web browser using WebGPU, which runs inference entirely on your device without uploading images to external servers.

The model's architecture builds on Liquid AI's previous releases, incorporating a SigLIP2 400-million-parameter vision encoder and training on approximately 34 trillion tokens. The company doubled the vocabulary size to 128,000 tokens to better support non-Latin scripts, and scaled vision pretraining by four times while improving data quality with synthetic image captions, optical character recognition (OCR) data, grounding annotations, and instruction-following examples.

The post-training process combined supervised fine-tuning with knowledge distillation from larger teacher models, followed by multi-reward reinforcement learning (RL), a technique where the model learns from multiple feedback signals simultaneously. This approach allowed Liquid AI to achieve strong performance without the computational expense of reasoning-based models that spend extra processing cycles thinking through problems step-by-step.

What Do These Results Mean for the AI Industry?

The LFM2.5-VL-3B's performance suggests that efficiency and capability are not mutually exclusive. Across 28 benchmarks covering multilingual visual comprehension, document understanding, object detection, and screen navigation, the model came within 0.7 percent of the larger 4.7-billion-parameter Qwen 3.5 model while using significantly less memory and power.

This efficiency matters beyond raw performance numbers. Smaller models reduce the environmental cost of AI deployment, lower infrastructure expenses for companies running AI services, and enable AI capabilities on devices where cloud connectivity is unavailable or undesirable. The model is available immediately on Hugging Face, a platform where researchers share open-source models, and through Liquid AI's playground for direct testing.

The shift toward efficient inference represents a maturation of the AI field. Early AI development prioritized raw capability, leading to models with hundreds of billions of parameters. As the field matures, the focus is shifting toward practical deployment, where speed, cost, and privacy matter as much as benchmark scores. The LFM2.5-VL-3B demonstrates that this transition is already underway.