Liquid AI shipped LFM2.5-VL-3B on August 12, 2026 — a 3.1B-parameter vision-language model for real-time, on-device inference. Open weights are on Hugging Face. The model runs in approximately 3 GB of memory and decodes at 228 tokens/second on an Apple M5 Max without a discrete GPU.
The architecture pairs a SigLIP2 400M NaFlex vision encoder with Liquid's LFM2.5-2.6B text backbone, pre-trained on roughly 34 trillion tokens — 4× more vision data than the prior release. The tokenizer expanded to 128K by in-place extension rather than full retraining, adding non-Latin script support. Post-training involved supervised fine-tuning with knowledge distillation, "Antidoom" training, and multi-reward reinforcement learning. The model skips chain-of-thought at inference and answers directly — a latency-first choice.
Four benchmark areas show the sharpest gains over LFM2-VL-3B. Screen and UI understanding jumped dramatically: ScreenSpot-v2 Desktop from 6.0 to 78.7, Mobile from 7.6 to 81.2, Web from 2.5 to 82.2. Grounding improved to 87.9 on RefCOCO-avg, ahead of Gemma-4-E4B (8B) at 72.1 despite LFM2.5-VL-3B's smaller size. ToolSandbox climbed from 26.4 to 59.5, now matching Gemma-4-E2B-it (56.5). Multi-image reasoning is now supported.
| Benchmark | LFM2-VL-3B (prev) | LFM2.5-VL-3B | Competitor (model) |
|---|---|---|---|
| ScreenSpot-v2 Desktop | 6.0 | 78.7 | — |
| ScreenSpot-v2 Mobile | 7.6 | 81.2 | — |
| ScreenSpot-v2 Web | 2.5 | 82.2 | — |
| RefCOCO-avg (grounding) | — | 87.9 | 72.1 · Gemma-4-E4B (8B) |
| ToolSandbox | 26.4 | 59.5 | 56.5 · Gemma-4-E2B-it |
Hardware performance matters for deployment. The M5 Max reaches 228 tokens/second, AMD Ryzen AI Max+ 395 reaches 116 tokens/second, and Samsung Galaxy S26 Ultra reaches 20 tokens/second fully on-device. On an H100 at high concurrency, Liquid reports 11,000 tokens/second output throughput — roughly 2× what 4B-class models produce and faster than 2B competitors in their testing. That translates to 1 billion output tokens per day on a single H100.
| Hardware Platform | Deployment Mode | Throughput (tokens/sec) |
|---|---|---|
| Apple M5 Max | On-device (no discrete GPU) | 228 |
| AMD Ryzen AI Max+ 395 | On-device | 116 |
| Samsung Galaxy S26 Ultra | Fully on-device mobile | 20 |
| NVIDIA H100 | Server, high concurrency | 11,000 |
Day-one inference support spans llama.cpp, MLX, vLLM, SGLang, and ONNX. MLX and llama.cpp are production-ready on Apple Silicon and AMD AI-accelerated laptops. vLLM and SGLang provide the throughput numbers above for server deployment. ONNX support enables mobile runtimes and hardware vendor SDKs without custom integration.
On overall vision tasks, LFM2.5-VL-3B averages 69.4, above Gemma-4-E4B (59.7) and Qwen3.5-4B (70.1) despite those models having more parameters. Document understanding: DocVQA 91.1, ChartQA 81.3. The model trails on MMMU (48.4 vs. 50.3) and OCRBench v2 English (47.5 vs. 58.7) — a trade-off reflecting its focus. Deep academic reasoning is not the target workload. The model prioritizes high-volume, low-latency vision pipelines.
| Model | Params | Overall VL avg | DocVQA | ChartQA | MMMU | OCRBench v2 (EN) |
|---|---|---|---|---|---|---|
| LFM2.5-VL-3B | 3.1 B | 69.4 | 91.1 | 81.3 | 48.4 | 47.5 |
| Qwen3.5-4B | 4 B | 70.1 | — | — | — | — |
| Gemma-4-E4B | 8 B | 59.7 | — | — | — | — |
The practical scope is narrow by design: screen automation agents, on-device document OCR, grounding-heavy workflows like visual search and object detection via natural language, and tool-calling agents that need vision context without cloud round-trip latency. For pipelines needing chain-of-thought or MMMU-class reasoning, look elsewhere. For those running on laptops, mobile devices, or single GPUs at high throughput, LFM2.5-VL-3B is currently one of the few sub-4B models with grounding and screen-understanding scores that justify the trade.