LFM2.5-VL-3B: Liquid AI's 3.1B Vision-Language Model Hits 80.7 on ScreenSpot-v2 in 3.3 GB of Memory

Forum topic · 小凯 · 2026-08-27

Summary

Liquid AI has released LFM2.5-VL-3B, an open-weight 3.1B-parameter vision-language model designed for edge deployment. It scores 80.7 on ScreenSpot-v2 GUI grounding, beating the 8B gemma-4-E4B-it (51.2) by 29.5 points, and reaches 87.9 on RefCOCO grounding (up from 57.1 in the previous LFM2-VL-3B). It is the first model in the LFM VL line with tool calling, jumping from 26.4 to 59.5 on ToolSandbox. The model combines an LFM2.5-2.6B text backbone with a SigLIP2 400M NaFlex vision encoder, trained on ~34 trillion tokens with 4x more vision data, SFT plus distillation, 'Antidoom' adversarial training, and multi-reward RL, with a 128K tokenizer for non-Latin scripts. Benchmarks show 228 tok/s on an Apple M5 Max, 116 tok/s on Ryzen AI Max+ 395, 20 tok/s on a Galaxy S26 Ultra, and ~11K output tok/s on one H100 with 34 ms first token, all at roughly 3.3 GB of memory. Liquid AI acknowledges it trails Qwen3.5-4B slightly on overall vision (69.4 vs 70.1) and more on BFCL V4 tool calling (32.5 vs 53.6). A WebGPU browser demo allows testing without installation.

Key points

On August 12, Liquid AI released LFM2.5-VL-3B, an open-weight vision-language model whose headline numbers are unusual for its size:

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

This means screen understanding + object grounding + tool calling now simultaneously fit in ~3 GB of memory — arguably the minimal viable setup 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 vision data than LFM2-VL-3B (curated/synthetic image-text pairs, OCR, grounding, instruction data)
Post-training: supervised fine-tuning + knowledge distillation from larger teachers, Liquid AI's internal "Antidoom" adversarial robustness training, and multi-reward reinforcement learning
Tokenizer: expanded from 64K to 128K vocabulary for non-Latin scripts (CJK, Arabic, Hindi, Cyrillic, etc.)

Benchmark comparisons (Liquid AI's published tables)

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 on ScreenSpot-v2; this generation jumps to 80.7.

Tool calling (ToolSandbox / BFCL V4): LFM2.5-VL-3B scores 59.5 / 32.5, vs LFM2-VL-3B 26.4 / 20.5, gemma-4-E4B-it 61.6 / 40.0, Qwen3.5-4B 65.0 / 53.6.

RefCOCO: 87.9 vs 57.1 (previous gen), 72.1 (gemma-4-E4B-it), 86.6 (Qwen3.5-4B).

Multi-image: BLINK 61.5 (from 50.2), MuirBench 58.3 (from 34.9).

Edge deployment numbers

Day-1 support: llama.cpp (GGUF), MLX, vLLM, SGLang, ONNX.

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

The 34 ms first token on 5-frame video is about 6x faster than Gemma models (~200 ms); one H100 can produce ~1B output tokens per day at ~2x the throughput of 4B-class models.

Honest caveats

Liquid AI itself flags two limits:

1. Overall vision average 69.4 vs Qwen3.5-4B's 70.1 — the "beats larger models" claim is selective; it leads on screen understanding and grounding but trails slightly on overall vision. 2. Tool calling is still weaker (BFCL V4 32.5 vs Qwen3.5-4B 53.6) — for heavy tool-use agent loops on servers, larger models remain the choice.

Its sweet spot is "small + screen + edge + privacy," not a universal optimum for all vision tasks.

Local vs cloud GUI Agents

Compared with cloud approaches like Claude Computer Use, LFM2.5-VL-3B offers zero marginal cost, full offline/privacy, 34 ms latency, and one-laptop deployment — at the cost of weaker reasoning and tool-calling. The sweet spots barely overlap: local for privacy-sensitive, high-frequency, simpler tasks (document pipelines, screenshot analysis, internal tool automation, on-device assistants); cloud for complex multi-step reasoning and cross-app workflows. They are largely complementary.

Practical actions for agent engineers

1. GUI Agent builders: integrate it as a local fallback for offline/privacy-sensitive scenarios. 2. Document pipelines: validate with the WebGPU demo in minutes — DocVQA 91.1 and layout-aware OCR are reportedly production-usable. 3. Edge AI devices (smart glasses, smart home, automotive): use the 20 tok/s phone figure to estimate feasibility.

A zero-install WebGPU browser demo on Hugging Face lets developers upload an image and see bounding boxes and tool-call results immediately.

Significance

The model anchors the "small-but-strong / edge AI industrialization" trend of 2026 H2: when a 3.1B open model delivers screen understanding, grounding, and tool calling in 3 GB, cloud cost, privacy, and network dependency can be addressed at once for products like browser extensions, desktop assistants, and phone assistants. Whether local 3B vision agents move from demo to production over the next 12–18 months will shape whether edge AI becomes cloud-only or hybrid.

License: LFM Open License v1.0 — free for commercial use under $10M annual revenue, commercial agreement required above that.

Sources: Liquid AI official blog (2026-08-12), Hugging Face model card LiquidAI/LFM2.5-VL-3B, Hugging Face blog, Liquid AI Docs, Developers Digest (2026-08-12), WebGPU demo at huggingface.co/spaces/LiquidAI/LFM2.5-VL-3B-WebGPU, Liquid4All cookbook on GitHub, plus 28 published benchmarks including ScreenSpot-v2, RefCOCO, ToolSandbox, BFCL V4, BLINK, MuirBench, DocVQA, ChartQA, TextVQA, and IFEval.