LFM2.5-VL-3B

LFM2.5-VL-3B

Quick answer: LFM2.5-VL-3B is Liquid AI's 3.1B-parameter on-device vision-language model, released August 12, 2026, built on the same pre-trained base as LFM2.5-2.6B with a SigLIP2 400M NaFlex vision encoder. It leads its size class on screen/UI understanding (80.7% ScreenSpot-v2), grounding (87.9% RefCOCO), and multi-image reasoning (61.5% BLINK, 58.3% MuirBench), while also posting a much stronger function-calling score than its predecessor (59.5% ToolSandbox, up from 26.4%). It decodes at 228 tokens/s on an Apple M5 Max under ~3 GB of memory. Open weights under Liquid's lfm1.0 license.

At a Glance

Where LFM2.5-VL-3B leads

• 80.7% average ScreenSpot-v2 (screen/UI grounding) — far ahead of Gemma-4-E4B-it (51.2%) and Qwen3.5-4B (78.5%), close behind the larger InternVL-3.5-4B (84.1%)
• 87.9% RefCOCO-avg grounding — a 30-point jump over its predecessor LFM2-VL-3B (57.1%)
• 59.5% ToolSandbox function calling — more than doubled from LFM2-VL-3B's 26.4%
• 228 tokens/s on Apple M5 Max, ~11K output tokens/s at high GPU concurrency on a single H100 — fastest in its comparison set
• Under ~3 GB memory footprint; runs on a phone (Galaxy S26 Ultra) at 20 tokens/s

Where it lags

• 32.5% BFCL V4 function calling — trails the larger Gemma-4-E4B-it (40.0%) and Qwen3.5-4B (53.6%)
• 48.4% MMMU (val) and 30.5% MMMU-Pro — knowledge-heavy STEM reasoning is not this model's strength relative to larger comparison models
• It's a non-reasoning model — answers directly rather than via chain-of-thought, trading some accuracy ceiling for lower latency

Best for: on-device / edge vision-language workloads — screen and UI understanding, document/OCR reading, visual grounding, and multi-image reasoning where local, private, real-time inference matters more than raw STEM knowledge.

What LFM2.5-VL-3B Is

LFM2.5-VL-3B is Liquid AI's latest vision-language model, extending its previous LFM2-VL-3B release with major improvements to screen/UI understanding, function calling, grounding, and multi-image input. It builds on the same pre-trained base as LFM2.5-2.6B (Liquid's text-only sibling model) and integrates a SigLIP2 400M NaFlex vision encoder, pre-trained on roughly 34 trillion tokens with a 4x increase in vision-pretraining tokens over its predecessor and a doubled 128K-token vocabulary to better support non-Latin scripts.

The model is a non-reasoning vision-language model — it answers directly rather than emitting a chain of thought, which keeps latency low for real-time and on-device applications such as live screen navigation or on-device document capture. Post-training combines supervised fine-tuning (including knowledge distillation from a larger teacher model and Liquid's "Antidoom" safety training) with multi-reward reinforcement learning.

On Benchgen's read, LFM2.5-VL-3B's standout results cluster around practical, agent-relevant vision tasks — reading and navigating digital screens, grounding objects in images for tool use, and reasoning across multiple images or video frames — rather than raw academic knowledge benchmarks, where larger comparison models (Gemma-4-E4B-it, Qwen3.5-4B) still hold an edge.

Specifications

Liquid AI's launch blog does not separately specify the vision-language context window; the model shares its pre-trained text base with LFM2.5-2.6B, which uses a 131,072-token context window.

Pricing

Open weights, self-hostable — no per-token API pricing. Available on Hugging Face, with day-one support for llama.cpp (GGUF), MLX (Apple Silicon), vLLM, SGLang, and ONNX. Designed for zero-marginal-cost local inference: decodes at 228 tokens/s on an Apple M5 Max and 116 tokens/s on an AMD Ryzen AI Max+ 395, under ~3 GB of memory, and reaches roughly 11,000 output tokens/s at high concurrency on a single NVIDIA H100 GPU.

Public Benchmark Scores

Scores are self-reported by Liquid AI in the official LFM2.5-VL-3B launch post, evaluated with vLLM 0.26 in non-reasoning mode using each model's recommended generation parameters.

Notable In-House / Unlisted Benchmark Results

Liquid AI's launch blog also reports scores on additional vision and multilingual benchmarks not yet tracked as standalone Benchgen leaderboard entries — mostly general-purpose academic VLM benchmarks that overlap heavily with the metrics already listed above, or whose scale couldn't be cleanly cross-verified against an existing Bengen page:

• MMStar — 63.3%
• MME — 73.1 (normalized 0–100; original scale 0–2800)
• SimpleVQA — 35.4%
• SEED-Bench (image) — 77.7%
• MMBench (dev EN v1.1) — 81.0%
• CountBenchQA — 87.3%
• MMMB (multilingual) — 83.0%
• Multilingual MMBench — 79.5%
• MM-IFEval (multimodal instruction following) — 60.6%
• LogicVista — 37.4%
• InfographicVQA (val) — 70.2%
• OCRBench v1 — 84.2 (normalized; original scale 0–1000)
• OCRBench v2 (En) — 47.5%
• TextVQA (val) — 84.3%
• POPE — 88.7 (F1 across adversarial, popular, and random subsets)
• Overall vision benchmark average (28 tasks) — 69.4%, within 0.7 points of the larger 4.7B-parameter Qwen3.5-4B (70.1%)

LFM2.5-VL-3B vs Alternatives

Despite being the smallest model in the comparison, LFM2.5-VL-3B leads on screen/UI grounding and visual grounding, while the larger Qwen3.5-4B keeps a clear edge on function calling (BFCL V4). All comparison scores are as reported by Liquid AI in the same launch benchmark table; see each model's own page for scores from other sources.

FAQ

Is LFM2.5-VL-3B open source? Yes — available on Hugging Face under Liquid AI's lfm1.0 license, downloadable, fine-tunable, and deployable without restrictions.

How big is LFM2.5-VL-3B? 3.1 billion total parameters, combining an LFM2.5 text backbone with a SigLIP2 400M NaFlex vision encoder.

What is LFM2.5-VL-3B best used for? Liquid AI positions it for on-device screen/UI understanding, document and OCR reading, visual grounding, and multi-image reasoning — real-time and on-device vision-language workloads rather than knowledge-heavy STEM tasks.

How fast is LFM2.5-VL-3B? 228 tokens/s on an Apple M5 Max and 116 tokens/s on an AMD Ryzen AI Max+ 395, under ~3 GB of memory — it also runs on a phone (Galaxy S26 Ultra) at about 20 tokens/s. On GPU, it reaches roughly 11,000 output tokens/s at high concurrency on a single NVIDIA H100.

Does LFM2.5-VL-3B support reasoning/chain-of-thought? No — it's a non-reasoning model that answers directly, which keeps latency low for real-time and on-device applications.

Benchmark scores sourced from Liquid AI's official LFM2.5-VL-3B launch announcement (liquid.ai/blog, Aug 12, 2026) and the LiquidAI/LFM2.5-VL-3B model card on Hugging Face. Comparison-model scores (Gemma 4 E2B/E4B, Qwen3.5-4B) are as reported in that same table.