Liquid AI released LFM2.5-VL-3B on August 12, 2026, a 3.1-billion-parameter open-weight vision-language model built to run on phones, laptops, and single GPUs rather than in a data center. The model, announced on the company’s blog and posted to Hugging Face with weights available immediately, extends last year’s LFM2-VL-3B with stronger screen understanding, object grounding, multi-image reasoning, and function calling.
The company positions the release as its most capable vision model for self-hosted hardware. In the release post, Liquid AI describes it as “our most capable vision-language model,” one that “delivers competitive vision performance against models twice its size, while running faster across a range of CPU and GPU deployments even compared with models that have fewer parameters.”
All benchmark and speed figures in this release are vendor-reported: Liquid AI ran the evaluations itself using vLLM 0.26.0, with every model in non-reasoning mode and prompted to answer directly. No independent evaluations of the new model exist yet, which is the expected state of play on release day, though recent Unite.AI coverage of an Amazon study that evaluated several of the same comparator models illustrates why independently measured transcription behavior can diverge from clean benchmark scores.
How LFM2.5-VL-3B Reads Screens, Documents, and Multiple Images
The model pairs a SigLIP2 400M NaFlex vision encoder from Google with the same pre-trained backbone as Liquid AI’s LFM2.5-2.6B text model, released on August 4, 2026. According to the model card, it was pre-trained on roughly 34 trillion tokens with four times more vision data than its predecessor, and its vocabulary doubled to 128,000 tokens by extending the existing tokenizer rather than retraining, a change aimed at non-Latin scripts. Post-training combines supervised fine-tuning with knowledge distillation from a larger teacher model, followed by multi-reward reinforcement learning.
Four capability areas define the upgrade over LFM2-VL-3B. On screen understanding, the model averages 80.7 across the desktop, mobile, and web splits of ScreenSpot-v2, up from single digits on the prior release and well ahead of the 8-billion-parameter Gemma-4-E4B at 50.9, though behind the larger InternVL 3.5 4B at 84.2. On grounding, where the model returns bounding boxes for objects described in natural language, RefCOCO precision jumps from 57.1 to 87.9, a gain of more than 30 points that Liquid AI attributes to scaled synthetic grounding data.
Function calling is new to the company’s vision line. On ToolSandbox, which measures tool use, the score more than doubles from 26.4 to 59.5, putting the 3.1B model on par with Gemma-4-E2B and ahead of Qwen3.5-2B. Multi-image reasoning also improves sharply, with BLINK rising from 50.2 to 61.5 and MuirBench from 34.9 to 58.3.
Across the full 28-benchmark vision suite, LFM2.5-VL-3B averages 69.4, a 12-point improvement over its predecessor’s 57.2 and within 0.7 points of the larger 4.7-billion-parameter Qwen3.5-4B. The average conceals real texture, though: the model trails its rivals on some general suites, and it actually loses ground on CountBenchQA, which drops from 92.2 to 87.3.
What the Model Deliberately Leaves Out
LFM2.5-VL-3B is a non-reasoning model. It answers directly rather than generating an intermediate chain of thought, a design choice Liquid AI frames as latency optimization: the model card recommends it for “single-turn, high-throughput, low-latency tasks,” naming near-real-time object detection for automotive applications, batch OCR of scanned documents, and on-device translation of menus and road signs as intended workloads.
The same card states plainly what the model is not for. It “is not recommended for long-context, reasoning-intensive tasks, such as visual web design, or answering highly technical questions about blueprints.” Context length is 32,768 tokens, and the card flags its layout-annotation OCR format as experimental, warning that it “may change, may be unreliable, and may not be trivial to parse.” These are the vendor’s own stated constraints, and they mark where the capability claims stop.
The speed numbers that anchor the release follow from that design. Liquid AI reports decode speeds of 228 tokens per second on an Apple M5 Max and 116 tokens per second on an AMD Ryzen AI Max+ 395, in about 3 GB of memory, and 20 tokens per second on a Galaxy S26 Ultra. On a single NVIDIA H100, the company measures time to first token at roughly 34 milliseconds on a five-frame video clip, against around 200 milliseconds for the Gemma models, and output throughput near 11,000 tokens per second at high concurrency, which it says adds up to almost a billion output tokens per day on one card.
LFM2.5-VL-3B by the Numbers
- 3.1 billion parameters; 34 trillion pre-training tokens; 32,768-token context
- 69.4 average across 28 vision benchmarks, versus 57.2 for LFM2-VL-3B
- 80.7 average on ScreenSpot-v2 screen understanding, up from 2.5 to 7.6 per split on the prior model
- 87.9 RefCOCO grounding precision, up from 57.1
- 59.5 on ToolSandbox function calling, up from 26.4
- 228 tokens/s decode on Apple M5 Max; 20 tokens/s on Galaxy S26 Ultra; about 3 GB memory footprint
- Roughly 11,000 output tokens/s at high concurrency on a single H100
What Ships With LFM2.5-VL-3B
The weights are downloadable now from Hugging Face under an open-weight license, with quantized exports in GGUF, ONNX, and MLX formats and day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX runtimes. A WebGPU demo runs the model entirely in the browser, and the company lists 16 supported languages including English, Arabic, Chinese, Japanese, and Hindi.
The release rounds out the LFM2.5 family Liquid AI has been assembling through the second half of 2026: the LFM2.5-2.6B text model on August 4, 2026, encoder variants for long-context CPU inference in late July 2026, and now the flagship vision tier, which follows the smaller LFM2.5-VL-1.6B and 450M variants. Fine-tuning notebooks and documentation ship alongside the weights, so the model can be adapted to specific grounding, OCR, or tool-calling workloads from the first day.
