LFM2.5-VL-3B is worth a look if you're building offline document scanners, in-car assistants, or screen-reading agents that can't depend on a network connection. On an Apple M5 Max it generates 228 tokens per second locally and scores 61.5 on multi-image reasoning, but teams needing long-document reasoning or a hosted chat API should look at a larger cloud model instead.
LFM2.5-VL-3B is Liquid AI's newest vision-language model, released August 12, 2026, built to run entirely on-device instead of in the cloud. It scores 87.9 on RefCOCO object grounding, up from 57.1 in the prior generation, and reads full documents, charts, and screens locally in about 3.3GB of memory.
Provider: Liquid AI · Family: LFM2.5
Context window: 32,768 tokens
Input modalities: text, image · Output: text, tool-calls
About LFM2.5-VL-3B
LFM2.5-VL-3B is a vision-language model built by Liquid AI, the MIT CSAIL spinout behind the Liquid Foundation Model (LFM) family, and released August 12, 2026. It pairs a SigLIP2 NaFlex vision encoder (roughly 400 million parameters) with the LFM2.5-2.6B language backbone for a combined 3.1 billion parameters, and is designed to be the most capable vision-language model Liquid AI ships that still runs on personal hardware rather than a GPU server. It is the direct successor to LFM2-VL-3B, released October 24, 2025, built on the same 3-billion-parameter scale but with further mid- and post-training. Against its own predecessor, LFM2.5-VL-3B posts large gains on every headline benchmark Liquid AI published: RefCOCO grounding (macro precision@1) rose from 57.1 to 87.9, ScreenSpot Desktop screen understanding rose from 6.0 to 78.7, multi-image reasoning on BLINK rose from 50.2 to 61.5, and ToolSandbox function-calling rose from 26.4 to 59.5. On document and chart understanding, the model scores 81.3 on ChartQA and 68.5 on MathVista, and 63.3 on MMStar for general multimodal reasoning. Liquid AI's own release blog compares it against similarly sized open vision-language models including Qwen3.5-2B and 4B, InternVL 3.5 2B and 4B, and Gemma-4-E2B-it and E4B-it, rather than against frontier cloud models like GPT-5 or Gemini, since it is not built to compete on general reasoning benchmarks like SWE-bench or GPQA. LFM2.5-VL-3B has a 32,768-token context window, expanded further than most edge VLMs but well short of the 128,000-token capability Liquid AI has quoted for the standalone LFM2.5-2.6B text backbone; the vision-language variant trades some of that headroom for image tokens. The tokenizer vocabulary was doubled to 128,000 tokens specifically to support non-Latin scripts, and the model handles single and multiple images, full documents, and charts in one context window rather than requiring a separate OCR pass. The model accepts text and images (single or multiple, at native resolution) and outputs text and tool calls; it does not take audio or video input. Its core capabilities are full-page OCR with layout annotation, natural-language object grounding and detection, multi-image comparison and reasoning, and function calling for agentic workflows. Liquid AI's own documentation recommends it for single-turn, high-throughput, low-latency tasks, and explicitly advises against long-context or reasoning-intensive uses such as visual web design or answering highly technical blueprint questions. LFM2.5-VL-3B is released under the LFM Open License v1.0 (LFM1.0), an Apache 2.0-based license with one added revenue-based condition on commercial use (see pricing below for the exact threshold). Organizations past that threshold need a paid commercial license from Liquid AI, priced around deployment scale and support needs rather than published per-token rates, since there is no metered API for this model. The model ships with day-one support across llama.cpp, MLX (Apple Silicon), vLLM, SGLang, and ONNX runtimes, plus Liquid AI's own LEAP SDK for fine-tuning and deployment, and a WebGPU browser demo that needs no install. On-device throughput measured by Liquid AI ranges from 228 tokens per second on an Apple M5 Max and 116 tokens per second on an AMD Ryzen AI Max+ 395 down to 20 tokens per second on a Galaxy S26 Ultra phone; on server hardware, a single NVIDIA H100 handles roughly 11,000 tokens per second at high concurrency with a time-to-first-token near 34 milliseconds. The full model fits in about 3.3GB of memory at bfloat16, small enough to run alongside other apps on a laptop or high-end phone. No listing was found for the model on AWS Bedrock, Google Vertex AI, Azure, Together AI, or Fireworks as of August 2026; deployment is self-hosted or through Liquid AI's own Apollo access layer. Liquid AI has not published a dedicated system card, training data cutoff date, or named red-teaming partners for LFM2.5-VL-3B, and the company has no public responsible-scaling policy. Its broader safety framing, published on its company blog, is white-box explainability: favoring model behavior that can be inspected over black-box shortcuts, and publishing a model card with benchmark tables, intended use, and stated limitations for every LFM release, including this one. LFM2.5-VL-3B fits teams building offline document scanners, in-car or on-device screen assistants, and agentic tools that need object grounding or function calling without a network round trip. It is a poor fit for long-document reasoning, open-ended chat, or any workload that needs a hosted, pay-per-token API: teams with those needs are better served by a frontier multimodal model such as GPT-5, Gemini, or Claude, or by a same-class open competitor like Qwen3.5-VL or InternVL if self-hosting is still the goal. Liquid AI has not disclosed the specific training data mix for LFM2.5-VL-3B beyond describing the vision-language mid- and post-training stage as using roughly four times more vision data than the LFM2-VL-3B predecessor. Because the model runs on-device by default, using it does not require sending prompts, images, or documents to Liquid AI's servers; Liquid AI has not published SOC 2, ISO 27001, HIPAA, or GDPR certifications for its products as of August 2026.
Pricing
LFM2.5-VL-3B carries no metered price per token; it's a free download under Liquid AI's open license as long as your organization's annual revenue stays below $10 million. Once you cross that line, Liquid AI sells a commercial license scoped to your deployment instead of charging by usage, and qualifying nonprofits never hit the cap at all.
Key Features
- Screen and UI Understanding: Reads desktop and mobile screens well enough to locate a specific button or field from a plain-language instruction, the single biggest capability jump over the LFM2-VL-3B predecessor on Liquid AI's own ScreenSpot Desktop benchmark.
- Object Grounding and Detection: Locates objects and elements directly from natural-language queries rather than fixed label sets, measured on the RefCOCO grounding benchmark against Liquid AI's prior generation.
- Full-Page OCR and Chart Reading: Extracts text with layout annotation and reads charts and diagrams, scoring 81.3 on ChartQA and 68.5 on MathVista.
- On-Device Function Calling: Calls tools and functions from text-and-image input for offline agent workflows, scoring 59.5 on the ToolSandbox benchmark versus 26.4 for the prior generation.
- Runs in About 3GB of Memory: The full 3.1-billion-parameter model runs in a memory footprint small enough for a laptop or high-end phone instead of a GPU server.
Pros
- Ships day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX, so it drops into most existing on-device inference stacks without custom porting.
- Free to use commercially at any scale under the license's revenue threshold, with no per-token billing once deployed.
- Throughput ranges widely by hardware: an AMD Ryzen AI Max+ chip handles roughly 116 tokens each second, while a single H100 pushes past 11,000 tokens a second under heavy concurrent load, fast enough for both consumer and server deployments.
Cons
- Its context window is a fraction of the 128,000 tokens Liquid AI quotes for the standalone LFM2.5-2.6B text backbone, and the vendor's own model card recommends against long-context or reasoning-heavy tasks.
- No hosted pay-per-token API from Liquid AI or a major cloud marketplace was found; teams that want managed inference must self-host or use Liquid AI's own Apollo platform.
- On low-end mobile silicon, throughput drops sharply: a Galaxy S26 Ultra runs it at just 20 tokens per second, a fraction of its laptop-class speed.
Benchmarks
- mmstar: 63.3
- chartqa: 81.3
- mathvista: 68.5
- screenspot v2: 80.7
- blink multi image: 61.5
- screenspot desktop: 78.7
- refcoco macro prec1: 87.9
- toolsandbox function calling: 59.5
Frequently Asked Questions
How much does LFM2.5-VL-3B cost per 1M tokens?
There is no per-token price: LFM2.5-VL-3B is a downloadable open-weights model licensed under the LFM Open License v1.0. It is free to use commercially for any organization under $10 million in annual revenue; above that threshold, Liquid AI requires a paid commercial license priced around deployment scale rather than tokens.
How does LFM2.5-VL-3B compare on benchmarks vs Qwen3.5-VL?
Liquid AI's own release benchmarks put LFM2.5-VL-3B up against similarly sized open models instead: the Qwen3.5, InternVL 3.5, and Gemma-4-E families at the 2B-4B scale. On MMStar, a general multimodal reasoning benchmark, LFM2.5-VL-3B scores 63.3; Liquid AI has not published a head-to-head comparison against larger cloud models such as GPT-5 or Gemini, since this model targets on-device use rather than that tier.
Is LFM2.5-VL-3B open source or proprietary?
LFM2.5-VL-3B is released under Liquid AI's LFM Open License (LFM1.0), which is Apache-2.0-based with one added condition: free commercial use only applies below the license's $10 million revenue ceiling, and there's no copyleft requirement on your own fine-tunes. Crossing that threshold requires purchasing a commercial license directly from Liquid AI, though qualifying nonprofits are exempt from the revenue cap entirely.
Does LFM2.5-VL-3B train on user data?
Because inference runs on the user's own device by default, LFM2.5-VL-3B does not send prompts, images, or documents back to Liquid AI's servers the way a hosted API model would. Liquid AI has not published a specific training data cutoff date or a dedicated system card for this model, and has not disclosed SOC 2, ISO 27001, or GDPR certification for its products.
Who is LFM2.5-VL-3B best for and who should avoid it?
LFM2.5-VL-3B is the right pick for offline document scanning, on-device screen reading inside a car or app, and agent tools that need grounding or function calling with zero network dependency. Teams needing long-document reasoning, open-ended chat, or a hosted pay-per-token API should choose a frontier multimodal model like GPT-5 or Gemini instead, or a same-class open competitor such as Qwen3.5-VL or InternVL for teams that still want to self-host.