Liquid AI just released LFM2.5-VL-3B, and it's the clearest signal yet that the edge-vs-cloud tradeoff is dissolving. This 3.1B parameter vision-language model outperforms larger 4B-class competitors on grounding, screens, and documents while decoding at 228 tokens/second on an M5 Max—and even hitting 20 tok/s on a Galaxy S26 Ultra.
The model pairs a SigLIP2 400M NaFlex vision encoder with Liquid's LFM2.5-2.6B text backbone, trained on 34T tokens with 4× more vision data than its predecessor. But the headline isn't the parameter count or training scale. It's that Liquid shipped this with day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX. That's not a research release. That's production infrastructure.
Four Capabilities That Matter
Liquid highlights four new capabilities, and they're not incremental. Screen/UI understanding is now strong across desktop, mobile, and web—crucial for agent workflows that need to parse tool interfaces. Grounding improved dramatically, enabling natural-language object detection without bounding-box annotations at inference time.
Multi-image reasoning got an upgrade, and function calling improved sharply. On ToolSandbox, LFM2.5-VL-3B scores 59.5% versus its predecessor's 26.4%. That's not tuning—that's architectural.
The post-training pipeline is worth noting: supervised fine-tuning uses knowledge distillation from a larger teacher plus "Antidoom training" (a term I haven't seen elsewhere, likely Liquid's internal jargon for safety or robustness techniques), followed by multi-reward RL. The vocabulary doubled to 128K tokens to handle non-Latin scripts, extended in place rather than retrained from scratch—smart engineering to preserve learned representations.
Benchmark Reality Check
Liquid claims LFM2.5-VL-3B "leads its size class on real-world image tasks." The numbers back it up, mostly. On RealWorldQA it scores 73.1% versus Qwen3.5-4B's 67.1% and InternVL 3.5 4B's 67.7%. That's a 3B model beating 4B-class competitors.
The grounding story is even sharper. RefCOCO-avg jumps from 57.1% to 87.9%—nearly matching InternVL 3.5 4B's 88.8%. For screen understanding, the gaps are wild: ScreenSpot-v2 Desktop goes from 6.0% to 78.7%. That prior score suggests the old model essentially couldn't do UI grounding at all.
Document and OCR benchmarks show maturity: 91.1% on DocVQA, 84.2% on OCRBench v1, 81.3% on ChartQA. These aren't research curiosities—they're production use cases.
But context matters. On MMMU (validation), InternVL 3.5 4B scores 60.7% versus LFM2.5-VL-3B's 48.4%. On SimpleVQA, Qwen3.5-4B hits 40.7% versus Liquid's 35.4%. Edge-optimized models make tradeoffs, and complex reasoning is where you see it.
Speed That Ships
The inference numbers are the quiet flex. 228 tok/s on M5 Max, 116 tok/s on Ryzen AI Max+ 395, and 20 tok/s on a mobile SoC. At 3GB memory footprint, this fits on devices people actually own.
On GPU, Liquid claims LFM2.5-VL-3B is "the fastest on output throughput" at high concurrency, reaching ~11K tokens/second on H100—"roughly 2× the larger 4B-class models." That scales to nearly 1B output tokens per day on a single card. For production workloads (think document processing pipelines, UI testing agents, real-time visual assistants), that cost-per-token delta compounds fast.
The post emphasizes multi-frame inputs, where lower memory bandwidth and faster decode matter more. Agent loops that process screen sequences or compare images aren't one-shot queries—they're sustained inference.
What "Edge-First" Actually Means
Most labs optimize for cloud, then quantize down. Liquid is designing for the edge and scaling up capabilities within that constraint. The choice to answer directly instead of chain-of-thought reasoning isn't a limitation—it's a design decision. Real-time and on-device apps can't wait for multi-step reasoning traces.
The function-calling improvements (59.5% on ToolSandbox, 32.5% on BFCL V4) put LFM2.5-VL-3B on par with Gemma-4-E2B and Qwen3.5-2B for tool use. That's table stakes for agentic workflows, and Liquid delivered it in a model that runs on a laptop.
The ecosystem support is the proof of intent. llama.cpp means community deployment paths. MLX means Apple Silicon optimization. vLLM and SGLang mean production serving stacks. ONNX means cross-platform inference. Shipping all five on day one signals Liquid isn't targeting researchers—they're targeting builders.
The Bigger Pattern
LFM2.5-VL-3B arrives in a context where OpenAI and Anthropic chase longer reasoning traces and DeepMind optimizes for AlphaFold-scale science problems. Liquid is betting that most AI value will be delivered on-device or at the edge, where latency, privacy, and cost-per-inference dominate.
The screen/UI understanding focus is strategic. Agents need to interact with existing software, not just answer questions. Grounding enables robots and AR systems. Multi-image reasoning powers document workflows and visual search. Function calling turns models into tool users.
These aren't academic benchmarks. They're primitives for products.
The competitive dynamic is interesting. Qwen3.5 and InternVL 3.5 are strong baselines, and Gemma-4 brings Google's infrastructure. But none of them prioritize edge inference the way Liquid does. The tradeoff is visible—LFM2.5-VL-3B sacrifices some reasoning depth for speed and deployability—but for many use cases, that's the right exchange.
What's Missing
The blog post doesn't discuss training cost or data provenance beyond "curated and synthetic image-caption, OCR, grounding, and instruction-following sets." We don't know what "Antidoom training" actually entails. The multi-reward RL specifics are vague.
Evaluation methodology is cleaner than most releases: vLLM 0.26.0, recommended generation parameters, non-reasoning mode everywhere. But we don't get error bars, statistical significance, or failure-mode analysis. The benchmarks are normalized to 0–100, which aids comparison but obscures absolute performance.
The model is available on Hugging Face today with a browser WebGPU demo. Fine-tuning tutorials are promised. That's good distribution, but we'll see how the community adopts it versus sticking with Qwen or InternVL.
The Edge Is the New Default
If you're building vision agents, document processors, or UI automation, LFM2.5-VL-3B is now the edge baseline. It's faster than larger models, ships with production tooling, and benchmarks competitively on real-world tasks.
The next six months will show whether edge-first design wins market share or remains a niche. But Liquid is making the right bet: most AI inference won't happen in datacenters. It'll happen on the devices people already own, and the models that ship there need to be fast, capable, and cheap to run.
LFM2.5-VL-3B is all three.