Summary: Liquid AI released LFM2.5-VL-450M on April 8, 2026, replacing its previous 450M vision-language model. The update adds direct object detection and function calling, expands pretraining from 10T to 28T tokens, and supports eight languages for image understanding. Edge latency tests include 233 ms for a 256×256 frame on NVIDIA Jetson Orin, though some knowledge benchmarks declined slightly.
Liquid AI put out LFM2.5-VL-450M on April 8, 2026 — a vision-language model in the 450M-parameter class taking over from LFM2-VL-450M. The company bills it as “structured visual intelligence, edge to cloud”: small enough to live on phones and developer boards, with output orderly enough to plug straight into an application.
What this generation adds is easy to summarize: two capabilities production systems keep asking for — object detection, which emits bounding boxes directly, and function calling. Multilingual image understanding now spans eight languages, and instruction-following scores climb considerably. Weights are up on Hugging Face, with trials through the LEAP platform and the public playground.
Pretraining Data: 10T Tokens Becomes 28T
Deepest down, the pretraining corpus grew: 28T tokens against the previous generation’s 10T, close to three times the data in the same parameter class. Post-training mixes preference optimization with reinforcement learning, aimed — as the company puts it — at multimodal behavior in production. Layer counts, the tokenizer, and similar internals stay undisclosed, but the philosophy is Liquid AI’s standing line: build around edge devices’ memory and latency budgets first, not scale first and handle deployment after.
For teams building on it, the takeaway is straightforward: nearly three times the pretraining work in an unchanged parameter count lifts quality per parameter — the only lever an edge model can usefully pull.
What It Learns to Do: Draw Boxes, Call Functions
Object detection simply wasn’t available in the previous LFM2-VL-450M. This generation scores 81.28 on RefCOCO-M and can produce a bounding box for any object named in a picture. Text-mode function calling lands at 21.08 on BFCLv4 — a modest figure, but a 450M-class model that consistently emits a usable tool-call structure has cleared the entry bar.
Eight languages are in scope for multilingual image understanding: Arabic, Chinese, French, German, Japanese, Korean, Portuguese, and Spanish. Boxes plus tool calls make the standard edge-agent loop in miniature — study the frame, find the target, call a local tool to act on it.
Latency on Real Edge Hardware
Latency was measured on the Q4_0-quantized build. An NVIDIA Jetson Orin turns a 256×256 frame around in 233 ms and a 512×512 frame in 242 ms — quick enough to cover every frame of a 4 FPS stream. The Samsung S25 Ultra phone takes 950 ms and 2.4 seconds for the two resolutions; the AMD Ryzen AI Max+ 395 takes 637 ms and 944 ms.
With figures like these, running a VLM on a phone stops being a demo and enters the usable band: image Q&A answering in seconds, on-device document scanning, and localized inspection help — none of it needs a cloud round trip anymore. A pipeline that keeps pace with 4 FPS can also judge every frame — drop it, flag it, act on it — rather than sampling sparsely and gambling that the frame that mattered comes through.
Where It Gains, Where It Slips
Set beside its predecessor, most benchmarks tilt upward: MMStar 43.00 against 40.87, RealWorldQA 58.43 against 52.03, MMBench (dev en) 60.91 against 56.27, MMVet 41.10 against 33.85, OCRBench 684 against 657. Instruction following climbs steepest — MM-IFEval from 33.09 to 45.00, IFEval from 51.75 to 61.16, Multi-IF from 26.21 to 34.63. Multilingual MMMB leaps from 54.29 to 68.09, CountBench from 47.64 to 73.31.
A few numbers travel the other way: MMMU eases down from 34.44 to 32.67, InfoVQA from 44.56 to 43.02. The shape — big instruction-following and structure gains, knowledge reasoning flat or a touch lower — is what you expect when a small model spends its post-training budget on behaving properly in pipelines rather than absorbing more facts. Handle the methodology with care too: vision scores came out of VLMEvalKit, while multilingual MMMB was graded on translations produced by GPT-4.1-mini, so cross-benchmark comparisons warrant some skepticism.
A Pragmatic Wager on Small Models
The frontier race of 2026 is scored in parameters and gigawatts; Liquid AI wagers the other way, fitting usability inside 450M parameters. It lands on the fault line we noted in our opening outlook for 2026: one cohort of vendors chases the frontier while another pushes models onto the device. For products that guard privacy, run offline, or live under tight latency budgets, local vision models only grow more appealing.
The path to production is complete: pull the weights from Hugging Face and host them yourself, or lean on the hosted LEAP platform, and the docs carry a worked fine-tuning example building a satellite-imagery VLM on custom data. For a team that owes itself an answer on whether an edge VLM cuts it, few pilots come cheaper.