Apple's M5 chip is establishing itself as a serious contender for running advanced artificial intelligence models directly on personal devices, with new benchmarks showing it can process complex vision tasks at speeds previously reserved for cloud servers. The latest evidence comes from Liquid AI's release of LFM2.5-VL-3B, a compact vision-language model that decodes 228 tokens per second on an Apple M5 Max while consuming only about 3 gigabytes of memory.

This development matters because it signals a shift in how AI is being deployed. Rather than sending images and questions to remote data centers, users can now run capable AI models entirely on their own hardware, keeping sensitive information private and eliminating network latency. The M5's performance here isn't just respectable; it's competitive with much larger systems.

What Makes the M5 Stand Out for AI Workloads?

The M5 Max's ability to handle LFM2.5-VL-3B demonstrates Apple's unified memory architecture, which allows the processor to share memory between its CPU (central processing unit) and GPU (graphics processing unit) cores. This design eliminates the bottleneck of copying data back and forth between separate memory pools, a common limitation in traditional computer architectures. For vision-language models, which juggle both image processing and text understanding simultaneously, this unified approach translates directly into speed.

When tested against competing hardware, the M5 Max's 228 tokens per second substantially outpaced an AMD Ryzen AI Max+ 395, which achieved 116 tokens per second on the same model. To put this in perspective, a token is roughly equivalent to a word or small piece of text; at 228 tokens per second, the M5 Max can generate roughly 13,680 words per minute, making it suitable for real-time interactive applications.

How to Deploy Vision-Language Models on Apple Silicon

• MLX Framework: Liquid AI provides optimized inference support through MLX, Apple's machine learning framework specifically designed for Apple Silicon, enabling developers to run models with minimal setup and maximum efficiency.
• GGUF Format Support: The model is available in GGUF format through llama.cpp, a lightweight inference engine that allows even non-technical users to run the model locally on their Mac without requiring specialized software or cloud accounts.
• Cross-Platform Compatibility: Beyond Apple hardware, the model supports ONNX (Open Neural Network Exchange) format, allowing deployment across diverse accelerators and operating systems for maximum flexibility.

The practical implications extend beyond raw speed. LFM2.5-VL-3B excels at understanding digital screens, documents, and user interfaces. In benchmarks measuring screen comprehension, the model scored 80.7 on ScreenSpot-v2, significantly outperforming Google's Gemma-4-E4B model, which scored 51.2 on the same test. This capability makes it particularly useful for automation tasks, accessibility features, and intelligent document processing.

Why On-Device AI Performance Matters for Users

The shift toward on-device AI processing addresses growing privacy concerns and practical limitations of cloud-dependent systems. When sensitive documents, medical images, or personal photos are processed locally on an M5 Mac, they never leave the user's device. This eliminates transmission delays and removes the need to trust third-party servers with confidential information.

The M5's performance also demonstrates that Apple's silicon strategy, which began with the M1 chip in 2020, continues to mature. Apple designed these processors specifically for the tasks Mac users perform daily, including increasingly sophisticated AI workloads. The unified memory architecture, which Apple pioneered in its consumer chips, is now proving essential for efficient AI inference.

Beyond the M5 Max, the model even runs on mobile devices. LFM2.5-VL-3B achieves 20 tokens per second on a Samsung Galaxy S26 Ultra, meaning users can run capable vision-language AI directly on their phones for tasks like reading documents, answering questions about images, or understanding what's on their screen.

What Capabilities Does This Model Unlock?

LFM2.5-VL-3B represents a meaningful step forward in practical AI capabilities. The model improved significantly in four key areas compared to its predecessor. Grounding performance, which measures the model's ability to locate specific objects in images and return their coordinates, jumped from 57.1 to 87.9 on the RefCOCO benchmark, a 30-point improvement. Function calling, which enables the model to decide when to use external tools, more than doubled from 26.4 to 59.5 on the ToolSandbox benchmark.

The model also strengthened its ability to process multiple images simultaneously. On the BLINK benchmark, which tests multi-image reasoning, performance climbed from 50.2 to 61.5. These improvements mean the model can now handle more complex, real-world scenarios where users ask questions spanning multiple images or documents.

Screen and user interface understanding represents another major strength. The model can analyze website layouts, identify clickable elements, and navigate digital interfaces to complete tasks. This capability opens possibilities for automated accessibility features, testing tools, and intelligent automation that understands the visual structure of applications.

How Does This Compare to Cloud-Based Alternatives?

While cloud-based vision models often deliver higher accuracy through larger parameter counts and more computational resources, they introduce latency and privacy tradeoffs. LFM2.5-VL-3B's 3.1 billion parameters make it compact enough for local deployment while maintaining competitive performance. On average across 28 different benchmarks, the model comes within 0.7 percent of the larger Qwen 3.5 4B model, which has 4.7 billion parameters.

The speed advantage of local processing is substantial. On a single NVIDIA H100 GPU, LFM2.5-VL-3B achieves the highest output throughput of any tested model at approximately 11,000 tokens per second under sustained load. This efficiency means fewer servers are needed to handle the same workload, reducing infrastructure costs and energy consumption.

For users and developers, the M5's performance suggests that the era of requiring cloud services for capable AI is ending. As Apple Silicon continues to evolve, on-device AI processing will likely become the default for many applications, with cloud services reserved for specialized tasks requiring massive computational resources or collaborative processing.