Liquid AI released LFM2.5-VL-3B on August 12, calling it its most capable vision-language model to date. It is a 3.1B-parameter, open-weight, non-reasoning model — it answers directly without a chain of thought, trading that away for low latency aimed at real-time, on-device applications. (Official blog: https://www.liquid.ai/blog/lfm2-5-vl-3b)

Four core upgrades, all aimed at letting an agent read your screen locally

Compared to the previous LFM2-VL-3B, the upgrade concentrates on four fronts:

  1. Screen/UI understanding: The model shows strong understanding of digital screens across mobile, web, and desktop. It averages 80.7 on ScreenSpot-v2, far ahead of the much larger Gemma-4-E4B (51.2), above Qwen 3.5 4B (78.5), and just behind InternVL-3.5-4B (84.1).
  2. Function calling: New to the VL line, tool use and function calling work for both text-only and vision-text inputs. ToolSandbox more than doubles from 26.4 to 59.5, and BFCL v4 climbs from 20.5 to 32.5 — on par with Gemma-4-E2B and ahead of Qwen3.5-2B.
  3. Grounding: By scaling synthetic grounding data, RefCOCO precision@1 jumps from 57.1 to 87.9, a 30-point gain.
  4. Multi-image input: BLINK improves from 50.2 to 61.5, and MuirBench from 34.9 to 58.3.

Training recipe: 34T tokens, a 128K vocabulary, and a SigLIP2 encoder

The architecture follows LFM2.5-VL-1.6B / 450M, builds on the just-released LFM2.5-2.6B text base, and integrates a SigLIP2 400M NaFlex encoder. It is pre-trained on ~34T tokens; the vocabulary was extended in place to 128K to better support non-Latin scripts; vision pretraining was scaled 4x by tokens. The post-training pipeline is SFT (including knowledge distillation from a larger teacher and Antidoom training) followed by multi-reward RL.

Averaged over 28 benchmarks, LFM2.5-VL-3B scores 69.4 — significantly ahead of the much larger Gemma-4-E4B (8B, 59.7) and within 0.7% of the larger 4.7B Qwen3.5-4B (70.1).

Speed: 20 tokens/s on a phone, ~11K tokens/s on one H100

On-device numbers: 228 tokens/s decoding on an Apple M5 Max, 116 tokens/s on an AMD Ryzen AI Max+ 395, within about 3GB of memory; roughly 20 tokens/s on a Galaxy S26 Ultra — a screen-reading vision model that runs entirely, privately, on your own phone. On the GPU side, the non-reasoning design keeps time-to-first-token low: about 34ms on a 5-frame video clip versus roughly 200ms for the Gemma models. Under sustained vLLM 0.26 load, it reaches the highest output throughput of any tested model, about 11K tokens per second at high concurrency — roughly 2x larger 4B-class models, nearly 1B output tokens per day on a single H100.

Commentary: the first actually-usable moment for on-device VLMs

My take: the value of this release is not the benchmark table but the engineering tipping point it marks for on-device visual agents. Three things 3B-class VLMs have historically done badly — reading screens, calling tools, and grounding objects — are exactly what a local automation agent needs most. LFM2.5-VL-3B pulls all three up at once, with open weights and day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX. When screen understanding and function calling fit inside 3GB of memory, the cloud agent is no longer the only option — privacy-sensitive use cases (operating personal devices, processing local documents) get a systematic open alternative for the first time. For developers, note that its comparison set is consistently Gemma and Qwen at the same tier: the edge multimodal track is moving from demo to selection.