LFM2.5-VL-3B: Liquid AI's New 3B Vision-Language Model for the Edge

The first thing to know about LFM2.5-VL-3B is that it shipped in two half-steps. The weights appeared on Hugging Face on August 11, 2026 — a full bf16 checkpoint, a model card with a benchmark table and a sixteen-language README, and no announcement attached. The card showed zero downloads. The next day, August 12, Liquid AI posted the official write-up, "LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge," and the quiet release became a real one. If you are reading this the same week, you are reading one of the earliest independent write-ups of a model that almost nobody outside Liquid's lab has run yet — so here is what is actually knowable from the repo, and what is not.

What LFM2.5-VL-3B is

LFM2.5-VL-3B is a vision-language model — image and text in, text out — built on Liquid AI's LFM2.5 text backbone. Concretely: the language model is the LFM2.5-2.6B, paired with a SigLIP2 NaFlex 400M vision encoder, for about 3.1 billion total parameters. It keeps the family's hybrid architecture of short-range gated convolution blocks interleaved with grouped-query attention, a 128,000-token vocabulary, and a 32,768-token context window. It supports sixteen languages (English, Arabic, Chinese, French, German, Hindi, Indonesian, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Thai, Vietnamese), ships in bfloat16, and is licensed under Liquid's lfm1.0 open license.

It is not a from-scratch model. The card describes it as a multimodal variant of LFM2.5 that builds on the earlier LFM2-VL-3B with further mid- and post-training. That lineage matters: the vision skills are layered on top of a text model already pre-trained on roughly 34 trillion tokens, so the language side is not a toy — it is the same family engine the text models use, with a vision encoder welded on and re-trained for vision work.

What it can do

Liquid's write-up calls out four improvements over the previous LFM2-VL-3B: screen and UI understanding, function calling, grounding, and multi-image input. In plain terms, that means:

• Grounding — point at objects with natural-language queries ("the stop sign," "the price column") and get a normalized location back.

• Screen understanding — answer questions about what is on screen and where, benchmarked on ScreenSpot-v2.

• OCR with layout annotation — full-page OCR that also returns document structure, with 23 layout labels (text, title, list, table, table caption, chart, equation, code, page headers and footers, and more) as normalized integer coordinates.

• Tool use — a four-step function-calling flow that takes tool definitions as JSON, emits Pythonic calls between tool-call tokens, and returns a final plain-text answer.

• Multi-image input — reason across several images in a single turn.

Liquid's own guidance on where it does not fit is unusually specific. The card recommends it for single-turn, high-throughput, low-latency tasks — near-real-time object detection in automotive, batch OCR of scanned documents, on-device menu and road-sign translation — and explicitly warns against long-context reasoning-heavy work, visual web design, and highly technical blueprint questions.

The numbers — all vendor-reported

Every figure in this section comes from Liquid AI's own model card and blog post. None of it has been independently reproduced, and until a third party runs the checkpoint you should treat these as claims, not facts. That label is doing real work here, because the record is genuinely strong on paper:

• Grounding — RefCOCO macro precision@1 of 87.9, up from 57.1 on the previous LFM2-VL-3B — a roughly thirty-point jump Liquid credits to the vision post-training.

• Screen understanding — ScreenSpot-v2 average of 80.7, which Liquid reports ahead of the 78.5 they attribute to Qwen3.5-4B.

• Tool use — ToolSandbox 59.5, more than double the 26.4 Liquid measured on LFM2-VL-3B; BFCLv4 32.5, up from 20.5.

• General vision reasoning — MME 73.1, MMStar 63.3, RealWorldQA 73.1, MMMB 83.0, ChartQA 81.3, MathVista 68.5, POPE 88.7.

• OCR — OCRBenchv2 (English subset) 47.5, up from 43.9.

• Hard multimodal reasoning — MMMU Pro 30.5, BLINK 61.5, MuirBench 58.3 — the weaker corner, as expected for a 3B model.

• Speed, vendor-run — 228 tokens/s on an Apple M5 Max, 116 tokens/s on an AMD Ryzen AI Max+ 395, and 20 tokens/s on a Galaxy S26 Ultra, all within roughly 3.3 GB of memory. On an H100 with vLLM, Liquid claims about 11K output tokens/s at high concurrency and a time-to-first-token of roughly 34 ms on a five-frame clip, which it attributes to the model answering directly instead of reasoning.

That is a lot of claims for one 3B model, and it is worth restating the caveat in the footer of the model's own card: these are Liquid's numbers. The gap between "scores well in a lab eval" and "holds up on your data" is exactly where a model this new usually gets tripped up.

What isn't known yet

The honest list of open questions:

• No independent benchmark — as of this writing LFM2.5-VL-3B appears on no third-party leaderboard. Every score above is vendor-reported.

• No hosted endpoint — no inference provider serves it yet. This is a self-host story, period.

• No usage data — zero downloads in its first day means no community feedback, no fine-tunes, no "I ran it on X and it broke at Y." The first real-world reports are still ahead of us.

• An experimental feature — the layout-annotation output is flagged by Liquid itself as "experimental" and "may change, may be unreliable, and may not be trivial to parse." Do not build your parser around it yet.

• Thin third-party coverage — the week-one press is mostly short news roundups. Nobody has done a deep independent test.

None of that means the model is bad. It means it is unproven in the way any model is unproven in the days after release.

How to run it yourself

For a model this young, the ecosystem story is unusually good. Format variants landed the same day as the weights: GGUF for llama.cpp, ONNX, and five MLX quantizations for Apple Silicon, alongside the native bf16 checkpoint for Transformers, vLLM, and SGLang. Liquid lists day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX, plus a WebGPU browser demo. Recommended sampling parameters are temperature 0.2, top_k 50, repetition penalty 1.0, with vision handled by the model's processor config. If you want to try it on a laptop tonight, the MLX or GGUF builds are the shortest path.

What to watch

Two things decide whether LFM2.5-VL-3B matters beyond the hype cycle. First, whether the grounding and screen-understanding numbers survive independent reproduction — 80.7 on ScreenSpot-v2 and 87.9 on RefCOCO are the two claims most worth checking, because those are the ones that would make this a genuinely disruptive edge model. Second, whether a hosted endpoint appears, because that changes the calculus from "interesting self-host project" to "shippable today." And that is precisely when a routing layer earns its keep: one API across 200+ models means the day Liquid or a provider stands up an endpoint, you add LFM2.5-VL-3B next to the models you already call without changing your integration, at the provider's list price with zero markup, and automatic failover lets you point real traffic at it while a proven model covers the calls it fumbles. Until then it is a promising self-host story — and for a model a week old, that is already more than most ships get.