LFM2.5-VL-3B: Optimizing Vision-Language Performance for Edge Devices

Liquid AI has released LFM2.5-VL-3B, a vision-language model targeting on-device and edge inference, according to the model's Hugging Face release. The 3B-parameter model pairs a SigLIP2 400M NaFlex vision encoder with the pre-trained backbone of the LFM2.5-2.6B text model and reports decoding throughput on Apple M5 Max, Ryzen AI Max+ 395, and Galaxy S26 Ultra hardware.

Architecture and Training Pipeline

LFM2.5-VL-3B is pre-trained on approximately 34T tokens with 4x more vision data than prior Liquid AI releases, drawn from curated and synthetic image-caption, OCR, grounding, and instruction-following corpora. To accommodate non-Latin scripts, the vocabulary was doubled to 128K by extending the existing tokenizer in place rather than retraining from scratch — a parameter-efficient choice that preserves the backbone's learned representations but leaves unaddressed how the extension behaves on scripts underrepresented in the original tokenization. Post-training proceeds in two stages: supervised fine-tuning with knowledge distillation from an unnamed larger teacher and a procedure the release labels "Antidoom training," followed by multi-reward reinforcement learning. The omission of the teacher's identity and the absence of ablation isolating each stage's contribution limit reproducibility.

Reported Benchmarks

Evaluation was conducted using vLLM 0.26.0 under non-reasoning decoding, with each model run using its own "recommended generation parameters when available." LFM2.5-VL-3B is reported to lead its size class on real-world image tasks and to read digital content — documents, charts, on-screen UI elements — at competitive levels. On tool use, the model is positioned as on par with Gemma-4-E2B and Qwen3.5-2B. The release does not publish raw scores, confidence intervals, or per-task breakdowns, so the magnitude of the lead cannot be independently assessed. Constraining evaluation to non-reasoning mode systematically disadvantages reasoning-tuned baselines and is a non-trivial confounder when interpreting "size class leadership."

On-Device and GPU Throughput

Reported inference speeds include 228 tokens/s decoding on an M5 Max, 116 tokens/s on a Ryzen AI Max+ 395, and 20 tokens/s on a Galaxy S26 Ultra, with the model fitting in roughly 3 GB of memory. On GPU infrastructure, LFM2.5-VL-3B is described as the fastest model tested on multi-frame inputs and on output throughput at high concurrency, reaching approximately 11K tokens per second — about 2x the throughput of larger 4B-class models and ahead of 2B-class baselines, aggregating to nearly 1B output tokens per day on a single H100. Day-one support spans llama.cpp, MLX, vLLM, SGLang, and ONNX, with transformers compatibility at version 5.0.0 or higher.

What to Verify Before Adoption

Three points warrant independent measurement. First, the underlying benchmark scores and per-task variance are not disclosed, making it impossible to confirm whether the size-class lead is statistically meaningful or driven by a narrow subset of tasks. Second, the claimed GPU throughput advantage over 2B-class models is atypical under standard scaling and would require controlled comparison under identical batch sizes, sequence lengths, and KV-cache configurations to validate. Third, the 3 GB mobile footprint and the specific quantization regime used to achieve it are not reported, which is material for any deployment targeting that memory envelope on devices with competing system pressure.