Liquid AI's 3B vision model hits 228 tokens/s on M5 Max, 11K/s on GPU

Liquid AI is back with LFM2.5-VL-3B, a compact vision-language model that punches well above its weight class. Built on the same backbone as their 2.6B text model, it pairs a SigLIP2 400M NaFlex vision encoder with roughly 34 trillion tokens of pre-training — including 4x more vision data than previous releases. The tokenizer got doubled to 128K vocabulary to handle non-Latin scripts without retraining from scratch.

The headline numbers are hard to ignore. On an M5 Max, this thing decodes 228 tokens per second. A Ryzen AI Max+ 395 manages 116 tokens/s. Even a Galaxy S26 Ultra chugs along at 20 tokens/s fully on-device, fitting into about 3 GB of memory. But the real flex is on the server side: at high concurrency, LFM2.5-VL-3B pushes roughly 11,000 output tokens per second on a single H100. That’s nearly double what 4B-class models achieve and ahead of even smaller 2B competitors — roughly a billion output tokens per day per GPU.

Liquid AI didn’t just optimize for speed. The model’s four headline improvements are screen and UI understanding, grounding via natural language queries, multi-image reasoning, and significantly stronger function calling across both text-only and vision-text scenarios. On tool use benchmarks, it’s competitive with Gemma-4-E2B and Qwen3.5-2B. Vision benchmarks show it leading its size class on real-world image tasks while handling documents, charts, and on-screen elements with surprising competence.

Post-training involved supervised fine-tuning with knowledge distillation from a larger teacher plus Antidoom training, followed by multi-reward reinforcement learning. The result is a model that looks genuinely useful for high-volume, on-device workloads — captioning, OCR, document parsing, UI automation — without requiring a data center. Day-one support ships for llama.cpp, MLX, vLLM, SGLang, and ONNX. The question now is whether developers will actually reach for a 3B vision model when so many workflows default to API calls. The speed numbers suggest they might.