Liquid AI releases LFM2.5-VL-3B, an open-weight vision-language model with a 30-point grounding jump

Liquid AI has released LFM2.5-VL-3B, a 3.1-billion-parameter open-weight vision-language model built to run on phones, laptops, and single GPUs, with the company describing it as "our most capable vision-language model."

What's new

In its August 12, 2026 announcement, Liquid AI writes: "LFM2.5-VL-3B builds on our previous LFM2-VL-3B, with significant improvements in screen understanding, grounding, function calling, and multi-image input."

The benchmark gains are substantial for a model this size. Grounding precision on RefCOCO improves "from 57.1 to 87.9, a 30-point gain" over the prior release. Tool-use performance "more than doubles from 26.4 to 59.5" on ToolSandbox, and function-calling accuracy on BFCL v4 climbs "from 20.5 to 32.5." On screen understanding, the model posts an 80.7 average on ScreenSpot-v2, which Liquid AI contrasts with Gemma-4-E4B's 51.2 on the same benchmark — a considerably larger model.

Speed is a central selling point: Liquid AI reports on-device decoding of 228 tokens per second on an Apple M5 Max, and roughly 11,000 tokens per second in GPU throughput at high concurrency. The model is a non-reasoning model that answers directly rather than working through extended chain-of-thought, which keeps latency down for real-time and on-device use cases.

The weights are available today on Hugging Face under an open license, with quantized exports already published in GGUF, ONNX, and MLX formats, alongside day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX runtime.

Context

LFM2.5-VL-3B is Liquid AI's third notable open-weight release in three weeks, following LFM2.5-Encoders (July 28) and the on-device agentic model LFM2.5-2.6B (August 4). Across all three, the company's positioning has stayed consistent: small, efficient models tuned to run directly on consumer hardware — phones, laptops, and single GPUs — rather than requiring data-center inference.

That strategy sits opposite the frontier-scale approach taken by the major chat labs, and instead competes more directly with other efficiency-focused open releases, including Google's Gemma line, which Liquid AI explicitly benchmarks against in this release.

Why it matters

The screen-understanding and grounding gains are the notable part here: strong screen and UI-element grounding is a prerequisite for computer-use and on-device agent applications, an area every major lab is racing to improve. A 30-point RefCOCO jump in a single release cycle, in a 3B model small enough to run locally on a laptop, is a meaningful efficiency data point for anyone building agentic tooling that needs to operate without a round trip to a hosted API.

It also reinforces a broader trend: the gap between "frontier" and "edge-deployable" model capability keeps narrowing, with labs like Liquid AI iterating fast enough to beat larger, more established open models like Gemma on specific tasks just months after their prior release.