Post
19
A 3B Vision Model Running at 13 Tokens/Second — on a Phone
AI is moving from the cloud into your pocket.
I just tested Liquid AI's newly released LFM2.5-VL-3B locally on a Samsung Galaxy S26. The Q8_0 model ran through llama.cpp with Vulkan acceleration on the Qualcomm Adreno 840 GPU. I uploaded a real webpage screenshot from another device and asked the model to describe it.
It worked—and generated at 13.11 tokens per second.
That's significant for a 3B-class vision-language model running entirely on a phone. In my testing, LFM2.5-VL-3B is the most capable local vision model I've run on the S26 so far.
Liquid AI is building specifically for this world. Its focus is edge AI: capable models designed to run locally on phones, laptops, vehicles and other resource-constrained hardware instead of depending entirely on cloud inference. Its LFM family targets low memory usage, fast inference and deployment across CPUs, GPUs and NPUs.
LFM2.5-VL-3B brings vision into that strategy. It is an open-weight multimodal model capable of image understanding, OCR, document extraction, visual question answering and spatial reasoning. Liquid provides the weights publicly, including an official GGUF release for llama.cpp.
This is where open-weight AI gets interesting. A phone can now hold the model, process the image and generate useful multimodal responses locally at interactive speed.
No cloud GPU. No remote inference API. The phone is the AI computer.[LFM2.5-VL-3B — Hugging Face]( LiquidAI/LFM2.5-VL-3B)
[LFM2.5-VL-3B-GGUF — Hugging Face]( LiquidAI/LFM2.5-VL-3B-GGUF)
AI is moving from the cloud into your pocket.
I just tested Liquid AI's newly released LFM2.5-VL-3B locally on a Samsung Galaxy S26. The Q8_0 model ran through llama.cpp with Vulkan acceleration on the Qualcomm Adreno 840 GPU. I uploaded a real webpage screenshot from another device and asked the model to describe it.
It worked—and generated at 13.11 tokens per second.
That's significant for a 3B-class vision-language model running entirely on a phone. In my testing, LFM2.5-VL-3B is the most capable local vision model I've run on the S26 so far.
Liquid AI is building specifically for this world. Its focus is edge AI: capable models designed to run locally on phones, laptops, vehicles and other resource-constrained hardware instead of depending entirely on cloud inference. Its LFM family targets low memory usage, fast inference and deployment across CPUs, GPUs and NPUs.
LFM2.5-VL-3B brings vision into that strategy. It is an open-weight multimodal model capable of image understanding, OCR, document extraction, visual question answering and spatial reasoning. Liquid provides the weights publicly, including an official GGUF release for llama.cpp.
This is where open-weight AI gets interesting. A phone can now hold the model, process the image and generate useful multimodal responses locally at interactive speed.
No cloud GPU. No remote inference API. The phone is the AI computer.[LFM2.5-VL-3B — Hugging Face]( LiquidAI/LFM2.5-VL-3B)
[LFM2.5-VL-3B-GGUF — Hugging Face]( LiquidAI/LFM2.5-VL-3B-GGUF)