Liquid AI's Vision Model Lives on a Phone Because Its Working Memory Stays Small

A vision-language model, software that can look at an image and answer questions about it, just shipped as open weights, runs entirely on a phone or laptop, and never sends the image to a server. Liquid AI released LFM2.5-VL-3B, a model with 3.1 billion parameters that decodes 228 tokens per second on an Apple M5 Max, 116 on an AMD Ryzen AI Max+ 395, and 20 tokens per second on a Galaxy S26 Ultra, all in roughly 3.3 GB of device memory. The framing in most re-reports, that a small model "outpaces" larger rivals, gets the mechanism wrong. What changed is the memory model, not the parameter count.

The standard Transformer, since the 2017 "Attention Is All You Need" paper, builds a key-value cache that grows with every token the model reads. Long context means long memory. Vision-language models, which feed both an image and a string of text into the model, push that cache harder than text-only models do. On a phone, the key-value cache is usually the wall.

Liquid's text backbone, LFM2.5-2.6B, uses gated short convolutions and grouped query attention, not a vanilla Transformer. It descends from Liquid Time-Constant (LTC) networks, a continuous-time architecture pioneered at MIT CSAIL by Ramin Hasani, who is now Liquid AI's CEO. The practical consequence described in the LFM2 technical report: the model's state stays roughly the same size no matter how long the input gets. Working memory is fixed, not growing. That is what makes a 3.1B-parameter vision-language model small enough to live in 3.3 GB.

The trade is explicit. Hugging Face's own model card recommends LFM2.5-VL-3B for "single-turn low-latency tasks," not long-context reasoning. The architectural choice that opens the door to on-device inference closes the door on the long-context workloads where the largest VLMs earn their compute. For developers building a privacy-respecting image-Q&A tool that never leaves the device, that is an acceptable envelope. For a 100,000-token document analysis pipeline, it is not.

The vision side carries its own provenance. Liquid pairs the text backbone with SigLIP2 NaFlex, a 400M-parameter image encoder, then exports the whole package in GGUF, MLX, ONNX, vLLM, and SGLang formats with a four-step function-calling protocol for tool use. Sixteen languages, 32,768-token context window, 128,000-token vocabulary. The release lands as a developer kit, not a consumer product: there is no Liquid AI app, no chat interface, no bundled assistant. The audience is teams that have to keep visual data, including medical scans, legal exhibits, and proprietary equipment photos, out of someone else's data center.

That audience is real. The privacy story for on-device VLMs is not "your phone cannot be hacked" (local software can be exploited too) but that the image never leaves the device, which removes an entire category of cloud-side breach and a category of vendor trust. For hospitals processing patient imaging under HIPAA, for law firms reviewing confidential exhibits, for manufacturers inspecting proprietary parts, the practical question is not whether the model is small but whether the image can be sent.

Every benchmark number in the launch is vendor-reported. Liquid AI's blog post shows tables where the 3.1B model "matches" rivals with roughly 50% more parameters across several vision-language benchmarks. No independent third-party replication exists at launch, and "matches" is not "outpaces." At the size and latency envelope the model targets, the fixed-state backbone holds its own against Transformer VLMs of similar and slightly larger scale, in benchmarks the vendor chose to publish.

The release widens the set of problems a developer can solve without a server. The on-device VLMs that existed before this launch were either much smaller and much less capable, or required cloud round-trips. LFM2.5-VL-3B sits in the middle: capable enough for single-turn visual question answering, document OCR, and offline image captioning; small enough to ship in a mobile binary; architecturally honest about its limits. The next checkpoint is whether the open-weights release draws third-party benchmarks at scale, which is what would turn "vendor says" into "field reports."