Liquid AI and edge efficiency with LFM2.5-VL-3B: claimed numbers, stated limits

On 12 August 2026 Liquid AI released LFM2.5-VL-3B, an open-weights vision-language model with 3.1 billion parameters designed to run on consumer devices: phones, laptops and single GPUs. Built on a SigLIP2 vision encoder and the text backbone of LFM2.5-2.6B, the text-only model published eight days earlier, it supports a 32,768-token context window and 16 languages, Italian among them. According to the official blog, the goal is local processing, reducing the reliance on data centres.

The performance figures come from the company alone and were picked up by the trade press: an average of 69.4 across 28 vision benchmarks, matching InternVL-3.5-4B (69.4) and sitting 0.7 points below Qwen3.5-4B (70.1), with 3.1 billion parameters against 4.7. Alongside gains in areas such as ToolSandbox there is a regression on CountBenchQA, down to 87.3 from the 92.2 of the previous version. Without independent third-party testing, these numbers remain company claims, not yet verified by anyone outside Liquid AI — MarkTechPost itself notes that it is reporting the vendor's data without validating it.

On efficiency, Liquid AI claims 228 tokens per second on an Apple M5 Max and a memory footprint of roughly 3 GB. It is worth stressing that the conditions of these tests were never spelled out, which makes the reproducibility of the results hard to judge. Distribution is under the LFM1.0 licence: MarkTechPost reports that it allows free commercial use up to 10 million dollars in annual revenue, a detail that does not appear explicitly in the official model card on Hugging Face, which simply names the licence type without stating any revenue threshold.

The technical documentation draws clear boundaries around the model's field of use, stating that it "is not recommended for long-context, reasoning-intensive tasks" — visual web design, for instance, or highly technical questions about engineering projects. It is tuned instead for single-turn, high-throughput work such as document OCR and on-device translation, where low latency and keeping data local are the real operational advantages.

— Olya Efficiency is the new currency in the language-model market, but generation speed alone is not enough to validate an architecture. It is good to see companies pushing towards edge computing with competitive models; yet as long as the benchmarks stay self-declared and the test set-ups are described vaguely, the line between a genuine leap in quality and a marketing optimisation remains a thin one. Independent verification is not an optional extra, it is a necessity.