TL;DR
The developers of LFM2.5-VL-3B have announced a 3.1B-parameter vision-language model designed for local devices and high-volume inference. They report gains in screen understanding, grounding, multi-image analysis and tool calling, but the benchmark and speed results have not been independently verified.
The developers of LFM2.5-VL-3B have announced a 3.1B-parameter vision-language model built to read documents and screens, locate objects, compare multiple images and call software tools while running on local hardware. The release targets real-time edge applications, where memory use, latency and privacy can matter as much as model accuracy.
LFM2.5-VL-3B combines a SigLIP2 400M NaFlex vision encoder with the pretrained backbone used by the LFM2.5-2.6B text model. Its developers said pretraining covered about 34 trillion tokens and used four times as much vision data as the prior model, including image-caption, optical character recognition, grounding and instruction-following material.
The model has a 128,000-token vocabulary, doubled to improve coverage of non-Latin writing systems. According to the announcement, post-training included supervised fine-tuning, knowledge distillation, Antidoom training and multi-reward reinforcement learning. The model produces direct answers instead of exposing a reasoning process, a design choice intended to keep responses fast.
In the developers’ evaluation, LFM2.5-VL-3B recorded a 69.4 average across vision benchmarks, compared with 57.2 for LFM2-VL-3B. It scored 91.1 on DocVQA, 87.9 on the RefCOCO grounding average and between 78.7 and 82.2 on ScreenSpot-v2 desktop, mobile and web tests. These are developer-reported results, generated with vLLM 0.26.0 in non-reasoning mode rather than an independent comparison.
LFM2.5-VL-3B: better, faster vision for the edge
A compact vision-language model built to read documents and screens, locate objects, compare multiple images and call software tools on local hardware. The promise is compelling—but the published performance claims remain developer-reported.
Four upgrades build on LFM2
The model combines a SigLIP2 400M NaFlex vision encoder with the pretrained LFM2.5-2.6B text backbone, targeting multimodal tasks where latency, memory and privacy matter.
Understand interfaces
Interpret desktop, mobile and web screens for interface assistance, accessibility and device-control workflows.
Locate what users describe
Connect natural-language requests to objects, regions and controls within documents, images and screens.
Compare visual inputs
Reason across several images for inspection, comparison, document review and sequential visual tasks.
Act through approved tools
Translate text-only or visual requests into software-tool calls while producing direct, low-latency answers.
Edge speed is the headline claim
Reported output throughput varies sharply by device. These figures are directional, not apples-to-apples guarantees for every prompt, quantization format or deployment.
The developers also report roughly 11,000 tokens per second on one H100 at high concurrency. Hardware configuration, precision, batch size, prompt mix and end-to-end application overhead can materially change real-world results.
Large reported gains, limited verification
The announcement used vLLM 0.26.0 in non-reasoning mode. No independent replication is included in the supplied source.
| Evaluation | LFM2.5-VL-3B | Earlier model | Reported change | Evidence status |
|---|---|---|---|---|
| Vision benchmark average | 69.4 | 57.2 | +12.2 points | ~Developer-run |
| DocVQA | 91.1 | Not supplied | Document strength | ~Developer-run |
| RefCOCO grounding average | 87.9 | Not supplied | Strong localization | ~Developer-run |
| ScreenSpot-v2 | 78.7–82.2 | Not supplied | Desktop, mobile and web | ~Developer-run |
| ToolSandbox | 59.5 | 26.4 | +33.1 points | ~Selected test |
| BFCL V4 | 32.5 | 20.5 | +12.0 points | ~Selected test |
Comparable tool-use performance was claimed against selected Gemma-4-E2B and Qwen3.5-2B results. InternVL 3.5 was excluded because it lacks function-calling support.
Promising release. Open questions.
A compact model can reduce network delay and data exposure, but local execution alone does not establish application-level privacy, reliability or safety.
Treat the numbers as claims to test—not settled facts.
The capability profile is well matched to edge assistants, document extraction, accessibility, industrial inspection and interface automation. Production decisions should wait for device-specific testing with representative images, prompts and tools.
From camera input to controlled action
A responsible edge deployment connects model capability to explicit validation, tool permissions and monitoring.
Can it run without cloud access?
The developers say yes. A tested quantized deployment reportedly fits in about 3 GB, although compatibility and speed vary by hardware and format.
Where could it be useful?
Document extraction, visual question answering, interface assistance, accessibility, device control and systems that call approved tools.
What runtimes are supported?
Reported support includes llama.cpp, MLX, vLLM, SGLang and ONNX, with Transformers support beginning at version 5.10.1.
What evidence comes next?
Independent consumer-device benchmarks, production tool-calling trials and evaluations across varied documents, interfaces and languages.
Local Vision AI Gets Faster
A compact model that can process screens, documents and camera images locally could support assistants, accessibility tools, device controls and industrial systems without sending every input to a remote server. Local processing may reduce network delays and data exposure, although actual privacy depends on the surrounding application.
The developers said a quantized deployment fits in about 3 GB of memory and reaches 228 output tokens per second on an M5 Max, 116 on a Ryzen AI Max+ 395 and 20 on a Galaxy S26 Ultra. They also reported about 11,000 tokens per second on a single H100 at high concurrency. Hardware configuration, quantization settings and workload details will affect real-world results.
AI vision language model for local devices
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Four Upgrades Build on LFM2
The new model follows LFM2-VL-3B and focuses on four areas: digital screen interpretation, natural-language object grounding, analysis across several images and function calling in both text-only and vision-text tasks.
Developer tests show tool-use gains over the earlier model, including 59.5 on ToolSandbox, up from 26.4, and 32.5 on BFCL V4, up from 20.5. The developers said its tool performance was comparable with Gemma-4-E2B and Qwen3.5-2B on the selected tests. InternVL 3.5 was not included in those comparisons because it lacks function-calling support.
“Our most capable vision-language model you can run on your own hardware.”
— LFM2.5-VL-3B developers, writing on Hugging Face
As an affiliate, we earn on qualifying purchases.
Independent Testing Is Still Missing
It is not yet clear how closely the reported benchmark scores and throughput will match independent tests or production workloads. The source does not provide full hardware settings, precision levels, power consumption, prompt distributions or end-to-end latency measurements for every comparison.
The announcement also does not establish how reliably the model handles unfamiliar interfaces, poor-quality images or safety-sensitive tool calls. Dataset composition is described by category, but the supplied material does not give enough detail to measure contamination, geographic coverage or performance across individual languages.
As an affiliate, we earn on qualifying purchases.
Developers Can Test Local Deployments
LFM2.5-VL-3B is reported to have support across llama.cpp, MLX, vLLM, SGLang and ONNX, with Transformers support beginning at version 5.10.1. The next evidence will come from independent evaluations on consumer devices, production tool-calling systems and varied document and interface workloads.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is LFM2.5-VL-3B?
It is a 3.1B-parameter vision-language model designed to process text and images, including documents, screens and multiple-image inputs.
Can the model run without cloud access?
The developers say it can run fully on-device and fit in about 3 GB of memory in the tested deployment. Device compatibility and speed will vary by format and hardware.
What changed from LFM2-VL-3B?
The release adds reported improvements in screen understanding, object grounding and multi-image analysis, alongside stronger function calling and broader support for non-Latin scripts.
Are the performance claims independently verified?
No independent verification is included in the supplied source. The published numbers are developer-run benchmark results using selected models, settings and non-reasoning prompts.
What applications could use the model?
Potential uses include document extraction, interface assistance and visual question answering, plus systems that identify on-screen objects and call approved software tools.
Source: Hugging Face
