Liquid AI LFM2.5-VL-3B Hits 228 Tokens/Sec on Apple M5 Max

Liquid AI released LFM2.5-VL-3B today, an open-weight 3.1-billion-parameter vision-language model optimized for high-speed edge deployment. The release pairs the company’s recent LFM2.5-2.6B language backbone with a SigLIP2 400M NaFlex vision encoder, replacing the previous LFM2-VL-3B iteration. The system is designed as a “non-reasoning” model that skips chain-of-thought processing to return direct answers with minimal latency.

Architecture and Inference Speed

The model was pre-trained on approximately 34 trillion tokens, utilizing four times more vision data than its predecessor. Liquid AI expanded the tokenizer vocabulary to 128,000 tokens to improve support for non-Latin scripts. The resulting weights fit under a 3.3 GB memory footprint, leaving ample headroom on standard consumer hardware when you configure local AI inference pipelines.

Raw speed is the primary optimization target. On an Apple M5 Max, LFM2.5-VL-3B achieves 228 tokens per second. Running on an AMD Ryzen AI Max+ 395 yields 116 tokens per second. The system supports a 32K token context window, allowing for sustained document analysis or multi-image processing in local memory.

Benchmark Capabilities

The addition of scaled synthetic grounding data drove significant accuracy improvements across spatial and interface tasks. The model achieved an 80.7 average on ScreenSpot-v2, surpassing Qwen 3.5 4B at 78.5 and Gemma-4-E4B at 51.2. RefCOCO grounding precision rose 30 points to 87.9.

LFM2.5-VL-3B also introduces function calling to Liquid AI’s vision line. The model scored 59.5 on ToolSandbox, up from 26.4 in previous builds, alongside a 32.5 on BFCL v4. Multi-image reasoning tasks saw similar gains, moving from 50.2 to 61.5 on BLINK and from 34.9 to 58.3 on MUIRBench.

Ecosystem Support

The open weights are available on Hugging Face in native, GGUF, MLX, and ONNX formats. This breadth of format support means developers can run the model directly via llama.cpp, mlx-vlm, and Ollama for edge workloads, or route to vLLM and SGLang for server environments. Fine-tuning is supported through LoRA using the TRL and Unsloth frameworks.

The hardware efficiency aligns with recent industry shifts toward running LLMs locally for privacy-constrained enterprise environments. Liquid AI recently partnered with MacPaw to integrate these models for macOS users, while parallel integrations target robotics and automotive use cases via Mercedes-Benz.

If you build vision-dependent edge applications, the combination of robust MLX and GGUF support with a sub-4 GB memory footprint makes LFM2.5-VL-3B a practical replacement for larger cloud-dependent vision models.