On April 8, 2026, Liquid AI released LFM2.5-VL-450M, a 450M-parameter vision-language model and the successor to LFM2-VL-450M. The company’s tagline for it is “structured visual intelligence, edge to cloud”: small enough to run on phones and developer boards, with outputs structured enough to wire directly into an application.
The update direction is clear. The model gains object detection (it emits bounding boxes) and function calling — two capabilities production systems actually ask for — extends multilingual image understanding to eight languages, and posts large instruction-following gains. Weights are on Hugging Face, with hosted access through the LEAP platform and a public playground.
From 10T to 28T Tokens
The most fundamental change is pretraining scale: 28T tokens, up from 10T for the previous generation — nearly triple the data at the same parameter class. Post-training combines preference optimization with reinforcement learning, which the company describes as targeting multimodal behavior in production. Internals like layer counts and the tokenizer are not disclosed, but the design philosophy carries over from Liquid AI’s earlier work: architect for the memory and latency budgets of edge devices first, rather than scaling up and worrying about deployment later.
For downstream teams the significance is simple: roughly three times the pretraining compute went into the same parameter count, which pushes quality-per-parameter up. That is the only direction an edge model can usefully move.
New Skills: Boxes and Function Calling
The previous LFM2-VL-450M had no object-detection support; this generation scores 81.28 on RefCOCO-M and can output bounding boxes for objects named in an image. Text-mode function calling scores 21.08 on BFCLv4 — not a headline number, but for a 450M-class model, reliably emitting a usable tool-call structure is the entry ticket.
Multilingual image understanding covers eight languages: Arabic, Chinese, French, German, Japanese, Korean, Portuguese, and Spanish. Combine boxes with tool calls and you get the canonical edge-agent loop: look at the frame, localize the target, invoke a local tool to act on it.
Edge Latency, Measured
Liquid AI published Q4_0-quantized latency figures. On an NVIDIA Jetson Orin: 233 ms per frame at 256×256 and 242 ms at 512×512 — fast enough to process every frame of a 4 FPS stream. On a Samsung S25 Ultra phone: 950 ms and 2.4 seconds respectively. On an AMD Ryzen AI Max+ 395: 637 ms and 944 ms.
Those numbers move on-device VLMs from demo territory into the usable band: second-response image Q&A, on-device document scanning, and localized inspection assistance no longer need a cloud round trip. Keeping up with a 4 FPS stream also means the pipeline can decide per frame — drop, flag, or act — instead of sampling sparingly and hoping the interesting frame survives.
Wins and Losses on the Board
Against its predecessor, most benchmarks rise: MMStar 43.00 versus 40.87, RealWorldQA 58.43 versus 52.03, MMBench (dev en) 60.91 versus 56.27, MMVet 41.10 versus 33.85, OCRBench 684 versus 657. Instruction following improves most sharply: MM-IFEval climbs from 33.09 to 45.00, IFEval from 51.75 to 61.16, and Multi-IF from 26.21 to 34.63. The multilingual benchmark MMMB jumps from 54.29 to 68.09, and CountBench from 47.64 to 73.31.
There are regressions: MMMU slips from 34.44 to 32.67, and InfoVQA from 44.56 to 43.02. The overall pattern — large instruction-following and structure gains, flat-to-slightly-down knowledge reasoning — is what you expect when a small model’s post-training budget goes toward behaving correctly in pipelines rather than knowing more facts. Read the methodology with care, too — vision scores were measured with VLMEvalKit, while the multilingual MMMB uses translations generated by GPT-4.1-mini, so cross-benchmark comparisons deserve some skepticism.
The Pragmatic Small-Model Bet
Most of 2026’s frontier race is measured in parameters and gigawatts; Liquid AI is betting the other way, packing usability into 450M parameters. It matches the split we flagged in our opening outlook for 2026: one cohort of vendors chases the frontier while another pushes models onto the device itself. For privacy-sensitive, offline-first, or latency-bound products, local vision models only get more attractive from here.
The landing path is complete enough to try: download weights from Hugging Face and self-host, or use the hosted LEAP platform; the docs include a worked fine-tuning example that builds a satellite-imagery VLM on custom data. If your team needs to answer whether an edge VLM is good enough yet, this is one of the cheapest pilots available.
Sources
- LFM2.5-VL-450M: Structured Visual Intelligence, Edge to Cloud — Liquid AI
- LiquidAI/LFM2.5-VL-450M — Hugging Face
- Satellite VLM fine-tuning example — Liquid AI Docs
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.