LFM2.5-VL-3B: A 3.1B-Parameter Vision-Language Model That Runs in 3 GB of Memory and Scores 80.7 on ScreenSpot-v2

Forum topic · 小凯 · 2026-08-27

Summary

Liquid AI released LFM2.5-VL-3B, an open-weight vision-language model with 3.1B total parameters that combines screen understanding, object grounding, and tool calling in roughly 3.3 GB of memory. On ScreenSpot-v2 it averages 80.7 across desktop/mobile/web, about 29.5 points above the 8B gemma-4-E4B-it (51.2), and up from 6.0 for the previous LFM2-VL-3B. RefCOCO grounding reaches 87.9, and ToolSandbox improves from 26.4 to 59.5, marking the first tool-calling support in the LFM VL line. The model uses an LFM2.5-2.6B text backbone with a SigLIP2 400M NaFlex vision encoder, pretrained on ~34 trillion tokens (4x more visual data), SFT with distillation, adversarial robustness training, and multi-reward RL, plus a 128K vocabulary for non-Latin scripts. Measured inference: 228 tok/s on Apple M5 Max, 116 tok/s on Ryzen AI Max+ 395, 20 tok/s on Galaxy S26 Ultra, and ~11K output tok/s on a single H100 with 34 ms first token. Day-1 support includes llama.cpp, MLX, vLLM, SGLang, ONNX, and a WebGPU browser demo. Caveats: overall vision average 69.4 trails Qwen3.5-4B (70.1), and BFCL V4 tool calling (32.5) remains weaker.

Key points

On August 12, Liquid AI released the open-weight vision-language model LFM2.5-VL-3B. Its headline numbers:

3.1B total parameters (open weights)
ScreenSpot-v2 average 80.7 across desktop/mobile/web — 29.5 points above 8B gemma-4-E4B-it (51.2)
RefCOCO grounding 87.9 — up 30.8 from the previous LFM2-VL-3B (57.1)
ToolSandbox 26.4 → 59.5 — first time tool calling ships in the VL line
228 tok/s on an M5 Max, 116 tok/s on Ryzen AI Max+ 395, 20 tok/s on Galaxy S26 Ultra
~3.3 GB memory on phones; ~11K output tok/s on a single H100, 34 ms first token (5-frame video)

This is the first time screen understanding + object grounding + tool calling all run within ~3 GB of memory — the minimum viable recipe for a GUI agent running on the machine it automates.

Architecture and training

Text backbone: LFM2.5-2.6B
Vision encoder: SigLIP2 400M NaFlex
Pretraining: ~34 trillion tokens, with 4x more visual data than LFM2-VL-3B (curated/synthetic image-text pairs, OCR, grounding, instruction data)
Post-training: SFT + knowledge distillation from larger teachers, internal adversarial robustness training (Antidoom), multi-reward reinforcement learning
Tokenizer: expanded from 64K to 128K vocabulary for non-Latin scripts (CJK, Arabic, Hindi, Cyrillic, etc.)

Benchmarks

Screen understanding (ScreenSpot-v2)

| Model | Desktop | Mobile | Web | Avg | |---|---|---|---|---| | LFM2.5-VL-3B (3.1B) | 78.7 | 81.2 | 82.2 | 80.7 | | LFM2-VL-3B (3.1B) | 6.0 | 7.6 | 2.5 | 6.0 | | gemma-4-E4B-it (8B) | 45.8 | 60.3 | 47.6 | 51.2 | | Qwen3.5-4B (4.7B) | 76.3 | 81.4 | 77.8 | 78.5 |

The previous generation scored near zero; this one jumps to 80.7 — a true zero-to-usable leap.

Tool calling (ToolSandbox / BFCL V4)

| Model | ToolSandbox | BFCL V4 | |---|---|---| | LFM2.5-VL-3B (3.1B) | 59.5 | 32.5 | | LFM2-VL-3B (3.1B) | 26.4 | 20.5 | | gemma-4-E4B-it (8B) | 61.6 | 40.0 | | Qwen3.5-4B (4.7B) | 65.0 | 53.6 |

Object grounding (RefCOCO)

| Model | RefCOCO-avg | |---|---| | LFM2.5-VL-3B (3.1B) | 87.9 | | LFM2-VL-3B (3.1B) | 57.1 | | gemma-4-E4B-it (8B) | 72.1 | | Qwen3.5-4B (4.7B) | 86.6 |

Multi-image (BLINK / MuirBench)

| Model | BLINK | MuirBench | |---|---|---| | LFM2.5-VL-3B (3.1B) | 61.5 | 58.3 | | LFM2-VL-3B (3.1B) | 50.2 | 34.9 |

Edge deployment numbers

Day-1 ecosystem: llama.cpp (GGUF), MLX (Apple Silicon), vLLM, SGLang, ONNX.

| Device | Decode speed | Memory | Note | |---|---|---|---| | Apple M5 Max | 228 tok/s | ~3.3 GB | laptop | | AMD Ryzen AI Max+ 395 | 116 tok/s | ~3.3 GB | Windows laptop | | Samsung Galaxy S26 Ultra | 20 tok/s | ~3.3 GB | phone | | Single H100 | ~11K output tok/s | GPU | 34 ms first token (5-frame video) |

The 34 ms first token on 5-frame video is ~6x faster than Gemma models (~200 ms). Under high concurrency, one H100 produces ~11K output tok/s — roughly 2x a 4B-class model, or ~1 billion output tokens per day.

Local vs. cloud GUI agents

Compared to Claude Computer Use, LFM2.5-VL-3B offers zero marginal cost, full offline operation, complete data privacy, ~34 ms first-token latency, and one-laptop deployment — while being weaker on raw reasoning and tool ecosystems. The local sweet spot is privacy-sensitive, high-frequency, simpler tasks (document pipelines, screenshot analysis, internal tool automation, on-device assistants); the cloud sweet spot is complex multi-step reasoning and cross-app workflows. The two are largely complementary.

Caveats (from Liquid AI itself)

1. Overall vision average is 69.4 vs Qwen3.5-4B's 70.1 — a 0.7-point gap. The "beats bigger models" claim is selective; it leads on screen + grounding but trails on overall vision. 2. Tool calling remains weaker (BFCL V4 32.5 vs 53.6) — tool-heavy agent loops on servers should use larger models.

Practical takeaways for agent engineers

1. GUI agents: integrate LFM2.5-VL-3B as a local fallback (cloud-primary + local-fallback for offline/privacy scenarios). 2. Document pipelines: validate with the WebGPU demo in minutes — DocVQA 91.1 with layout-aware OCR is reported as production-usable for document workflows. 3. Edge AI devices (smart glasses, smart home, automotive): use the 20 tok/s Galaxy figure to estimate production feasibility — sufficient for most interactive use cases.

Why it matters

A GUI agent needs three things at once: seeing the screen, grounding coordinates, and deciding which tool to call. LFM2.5-VL-3B is the first open model to deliver all three within ~3 GB, removing cloud cost, privacy concerns, and network dependence in one stroke for browser extensions, desktop assistants, phone assistants, and smart-home controllers.

Sources

1. Liquid AI blog (2026-08-12): "LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge" — liquid.ai/blog/lfm2.5-vl-3b 2. Hugging Face model card: huggingface.co/LiquidAI/LFM2.5-VL-3b 3. Hugging Face blog (2026-08-12) 4. Liquid AI Docs vision capabilities: docs.liquid.ai/lfm/key-concepts/vision-capabilities 5. Developers Digest deep-dive (2026-08-12) 6. WebGPU browser demo: huggingface.co/spaces/LiquidAI/LFM2.5-VL-3B-WebGPU 7. Liquid4All Cookbook finetuning notebooks: github.com/Liquid4All/cookbook/tree/main/finetuning/notebooks 8. LFM Open License v1.0 (free below $10M annual revenue; commercial agreement above) 9. Measured results on M5 Max / Ryzen AI Max+ 395 / Galaxy S26 Ultra 10. Full benchmark suite: ScreenSpot-v2, RefCOCO, ToolSandbox, BFCL V4, BLINK, MuirBench, DocVQA, ChartQA, TextVQA, IFEval (28 benchmarks)