Liquid AI has released LFM2.5VL450M which is an update to its previous LFM2-VL450M model. The new release introduces bounding box prediction, improved instruction following, expanded multilingual understanding, and function calling support — all within a 450M-parameter footprint designed to run directly on edge hardware ranging from embedded AI modules like NVIDIA Jetson Orin, to mini-PC APUs like AMD Ryzen AI Max+ 395, to flagship phone SoCs like the Snapdragon 8 Elite inside the Samsung S25 Ultra.
What Is a Vision Language Model, and Why Size is Important?
To understand the vision-language models (VLM), it is helpful to first define them. A VLM is a model that can process both images and text together — you can send it a photo and ask questions about it in natural language, and it will respond. To run, most large VLMs need a lot of GPU memory. It’s problematic for deployments in the real world, such as with warehouse robots and smart glasses or retail shelf camera systems, which have limited compute power.
Liquid AI has responded to the constraint by developing a small model that can fit onto edge hardware, while supporting an important set of language and vision capabilities.
Training in Architecture
LFM2.5VL450M utilizes LFM2.5 350M as the language model and SigLIP2 86M shape-optimized SigLIP2 as vision encoder. This context window contains 32,768 tokens, with 65,536 words in the vocabulary.
For image handling, the model supports native resolution processing up to 512×512 pixels without upscaling, preserves non-standard aspect ratios without distortion, and uses a tiling strategy that splits large images into non-overlapping 512×512 patches while including thumbnail encoding for global context. It is crucial to encode thumbnails, as without them, the tiling strategy would only give local patches and no overall sense of the scene. The maximum number of image tokens, and the tile count can be adjusted at inference time to optimize performance without needing to retrain.
Liquid AI recommends the following generation parameters: temperature=0.1, min_p=0.15” repetition_penalty=1.05 Text is available in English and min_image_tokens=32, max_image_tokens=256” do_image_splitting=True Vision inputs.
Liquid AI compared the LFM2-VL450M to a scale of pre-training between 10T and 28T tokens, then used reinforcement learning and preference optimization to post-train to improve grounding and instruction following.
New Capabilities Over LFM2-VL-450M
Most notable is the addition of bounding box predictions. LFM2.5VL450M achieved 81.28 out of 100 on RefCOCOM. This is an improvement from the zero score on previous models. RefCOCO M is a visual benchmark which measures the accuracy of a model in locating an object given natural language descriptions. In practice, the model outputs structured JSON with normalized coordinates identifying where objects are in a scene — not just describing what is there, but also locating it. It is different than pure captioning of images and allows the model to be directly used with pipelines that require spatial outputs.
Also, multilingual support improved significantly. MMMB score improved from 54.29 up to 68.09 covering Arabic, Chinese Japanese Korean Portuguese and Spanish. The MMMB score has improved from 54.29 to 68.09, covering Arabic, Chinese and French as well as Japanese, Korean, Portuguese and Spanish.
The following of instructions also improved. MM-IFEval scores went from 32.93 to 45.00, meaning the model more reliably adheres to explicit constraints given in a prompt — for example, responding in a particular format or restricting output to specific fields.
The previous model didn’t have this capability. Function calling allows the model to be used in agentic pipelines where it needs to invoke external tools — for instance, calling a weather API or triggering an action in a downstream system.
Benchmark Performance
For most vision tasks, the LFM2.5VL450M is superior to both SmolVLM2-500M and LFM2-VL450M. Other notable scores are 86.93 for POPE, 684 for OCRBench and 60.91 for MMBench(dev en).
Beyond the headline figures, two benchmark improvements stand out. MMVet — which tests more open-ended visual understanding — improved from 33.85 to 41.10, a substantial relative gain. CountBench measures the ability of the model to count items in a given scene. The improvement from 47.64 – 73.31 is among the most significant improvements. InfoVQA was essentially flat at 43.02, compared to 44.56 in the prior model.
IFEval’s performance improved by 51.75 points to 61.16, and Multi-IF went from 26.21 points to 34.63. The model does not outperform on all tasks — MMMU (val) dropped slightly from 34.44 to 32.67 — and Liquid AI notes the model is not well-suited for knowledge-intensive tasks or fine-grained OCR.
Edge Inference performance
This LFM2.5-VL450M Quantization with Q4_0 is compatible across all hardware targets, including embedded AI modules such as Jetson Orin and mini-PC APUs Ryzen AI Max+ 395, to smartphone SoCs Snapdragon 8 elite.
These numbers are telling us a lot. On Jetson Orin, the model processes a 256×256 image in 233ms and a 512×512 image in 242ms — staying well under 250ms at both resolutions. It is fast enough to handle every frame of a four-frame video stream, with complete vision language understanding and not only detection. On Samsung S25 Ultra, latency is 950ms for 256×256 and 2.4 seconds for 512×512. On AMD Ryzen AI Max+ 395, it is 637ms for 256×256 and 944ms for 512×512 — under one second for the smaller resolution on both consumer devices, which keeps interactive applications responsive.
Real World Use Cases
LFM2.5VL-450M excels in real-world deployments, where low-latency, compact, structured outputs and semantic reasoning are most important. This includes settings that demand offline processing or device-based processing for privacy.
A common limitation in industrial automation occurs when computing resources are limited. This is the case for environments like passenger cars, agricultural equipment, and warehouses. Perception models can only be based on bounding boxes. LFM2.5-VL-450M goes further, providing grounded scene understanding in a single pass — enabling richer outputs for settings like warehouse aisles, including worker actions, forklift movement, and inventory flow — while still fitting existing edge hardware like a Jetson Orin.
Wearables such as dashcams, smart glasses, body worn assistants, security monitors and industrial monitors that are always on cannot support large cloud streams or perception stacks. A VLM that is efficient can generate compact semantic outputs, transforming raw video into structured understanding. This allows for low compute requirements and privacy to be maintained.
The cost of deploying richer visual reasoning at large scale is prohibitive. Structured visual reasoning is now possible for these workloads with the LFM2.5-VL450M.
What you need to know
- LFM2.5 adds bounding boxes for the first., scoring 81.28 on RefCOCO-M versus zero on the previous model, enabling the model to output structured spatial coordinates for detected objects — not just describe what it sees.
- The pre-training scale was increased from 10T up to 28T tokensThe combination of post-training through preference optimization, reinforcement learning and other methods, results in consistent benchmark gains for both language and vision tasks on the LFM2-VL450M.
- This model is run on cutting-edge hardware, with a sub-250ms latency., processing a 512×512 image in 242ms on NVIDIA Jetson Orin with Q4_0 quantization — fast enough for full vision-language understanding on every frame of a 4 FPS video stream without cloud offloading.
- Visual understanding of multilingualism has significantly improvedMMMB’s scores have increased from 54.29 points to 68.09 points across Arabic, Chinese (including Japanese), French, German and Korean.
Check out the Technical details You can also find out more about the following: Model Weight. Also, feel free to follow us on Twitter Don’t forget about our 120k+ ML SubReddit Subscribe Now our Newsletter. Wait! What? now you can join us on telegram as well.
You can partner with us to promote your GitHub Repository OR Hugging Page OR New Product Launch OR Webinar, etc.? Connect with us
