LFM2.5-VL-450M: Liquid AI's 450M vision model that runs in a browser

What is LFM2.5-VL-450M

Most vision-language models compete on scale — billions of parameters, hundreds of GPU-hours for inference. Liquid AI takes the opposite approach. LFM2.5-VL-450M is a 450M-parameter multimodal model that understands images and text across 9 languages, predicts bounding boxes, supports function calling, and runs real-time video captioning directly in a browser via WebGPU.

At 0.4B parameters, it is roughly 500x smaller than frontier models like MiniMax-M2.7 or GPT-5. Yet it outperforms SmolVLM2-500M on nearly every benchmark and introduces capabilities — visual grounding, tool use — that models 10x its size often lack.

---

Architecture

LFM2.5-VL-450M combines two components:

• LFM2.5-350M — a 350M dense Transformer serving as the language backbone (32K context window, 65K vocabulary)
• SigLIP2 NaFlex — an 86M shape-optimized vision encoder that processes images at native resolution up to 512×512 without upscaling or aspect ratio distortion

For larger images, the encoder uses an adaptive tiling strategy: non-overlapping 512×512 patches with a thumbnail encoding for global context. Users can tune the quality/speed tradeoff at inference time by adjusting max_image_tokens (32–256) and tile count — no retraining required.

The model ships in BF16 (safetensors), with GGUF, ONNX, and multiple MLX quantizations (4-bit through BF16) for Apple Silicon.

---

Benchmark results

Vision understanding

Benchmark	LFM2.5-VL-450M	LFM2-VL-450M	SmolVLM2-500M
MMStar	43.0	40.9	38.2
RealWorldQA	58.4	52.0	49.9
MMBench (dev en)	60.9	56.3	52.3
POPE	86.9	83.8	82.7
MMVet	41.1	33.9	29.9
OCRBench	684	657	609
MM-IFEval	45.0	33.1	11.3
CountBench	73.3	47.6	61.8
RefCOCO-M	81.3	—	—

LFM2.5-VL leads on every vision benchmark except MMMU (32.7 vs SmolVLM2’s 34.1 — a knowledge-intensive benchmark where larger models have a structural advantage). The MM-IFEval jump from 33.1 to 45.0 reflects significantly better instruction following on visual tasks.

RefCOCO-M at 81.3 is a new capability: bounding box prediction for visual grounding, not available in the previous LFM2 generation.

Multilingual vision (MMMB)

LFM2.5-VL-450M	LFM2-VL-450M	SmolVLM2-500M
68.1	54.3	46.8

MMMB averages vision understanding across 8 languages (Arabic, Chinese, French, German, Japanese, Korean, Portuguese, Spanish). The +13.8 point improvement over the previous version is the largest single-benchmark gain.

Language and tool use

Benchmark	LFM2.5-VL-450M	LFM2-VL-450M	SmolVLM2-500M
MMLU Pro	19.3	17.2	13.6
IFEval	61.2	51.8	30.1
Multi-IF	34.6	26.2	6.8
BFCLv4	21.1	—	—

IFEval at 61.2 — double SmolVLM2’s score — shows that instruction following scales with training quality, not just parameter count. BFCLv4 is a new function calling benchmark; LFM2-VL did not support tool use at all.

---

Key capabilities

Visual grounding — M2.5-VL can predict bounding boxes in normalized [0,1] coordinates, returned as JSON arrays. This enables object detection workflows without a separate detection model.

Function calling — text-only tool use in ChatML format with <|tool_call_start|> / <|tool_call_end|> tokens. The model can decide when to call functions and format arguments correctly.

Multilingual vision — 9-language support is not bolted on: the model was trained with multilingual vision understanding as a first-class objective, scoring 68.1 on MMMB (vs 46.8 for SmolVLM2).

Inference-time flexibility — min_image_tokens and max_image_tokens let users trade quality for speed without retraining. A mobile deployment can use 32 tokens per image; a desktop pipeline can use 256.

---

Deployment

LFM2.5-VL-450M is designed to run everywhere:

Framework	Use case
Transformers	Simple inference, fine-tuning
vLLM	High-throughput GPU production
SGLang	High-throughput GPU production
llama.cpp	CPU inference, local deployment
ONNX Runtime	Cross-platform, hardware-accelerated
MLX	Apple Silicon (4-bit through BF16)
WebGPU	Browser-based, real-time video captioning

The WebGPU demo runs real-time video stream captioning entirely in the browser — no server, no API, no GPU required.

Default generation parameters: temperature=0.1, min_p=0.15, repetition_penalty=1.05.

Fine-tuning is supported via LoRA with both Unsloth and TRL.

---

Limitations

At 450M parameters, LFM2.5-VL is not suited for knowledge-intensive tasks — MMMU at 32.7 and MMLU Pro at 19.3 confirm this. It is a perception and instruction-following model, not a reasoning one.

Fine-grained OCR is acknowledged as a limitation despite the OCRBench score of 684. Function calling is text-only — tool use does not support vision input. Image processing is capped at 512×512 per tile, with larger images split into patches.

---

Conclusion

LFM2.5-VL-450M proves that useful vision-language capabilities do not require billions of parameters. Visual grounding, function calling, 9-language support, and real-time browser inference — all in a model that fits in 900MB of VRAM — makes this the most deployment-flexible VLM available today.

For edge applications, mobile deployments, browser-based tools, or any scenario where a 70B model is impractical, LFM2.5-VL fills a gap that larger models cannot reach by design.

Model: LiquidAI/LFM2.5-VL-450M · Paper: arxiv.org/abs/2511.23404