Running Multimodal on 1GB of RAM: Hands-On with LFM3 — Is It Enough?

Liquid AI has released LFM3, the next generation of Liquid Foundation Models. It's live on LM Studio at the same time — the model page says: the smallest model needs about 1GB of memory.

This release is a family, not a single model. Six instances covering text, vision, audio, omni, and decision-making, with parameters from 1.7B to 2.2B. Liquid AI's claim: a tri-hybrid architecture, new multimodal encoders, designed for specialized intelligence.

The text model LFM3-1.7B has a 256K context. On WebArena-Verified and AutomationBench, the official numbers beat Qwen3.5-9B. A 128K-token KV cache takes 1.6GB, while Qwen3.5-4B needs 4.3GB. In long conversations, the smaller model actually leaves more memory for the application. Liquid AI also ran more than 300,000 agents in parallel to analyze the Enron email archive.

The vision model LFM3-VL-2B supports video input. It averages 71.4 across 62 vision benchmarks; the company says it matches Gemma 4 12B, which is 6x larger. It has only 1.92B parameters. It also brings spatial reasoning — locating objects and understanding their relative positions — which the company says provides a foundation for robotics applications.

The audio model LFM3-Audio-2B takes text or audio input and outputs text, speech, or both. 18 languages, with support for spoken tool calls. The official internal evaluation is 58.5, versus 44.3 for Qwen2.5-Omni-3B. The LiquidCodec tokenizer synthesizes speech at roughly 87x real time single-threaded on an Apple M4 Max. A causal encoder can process while listening, cutting down the wait.

The Omni model LFM3-Omni-2B handles vision, audio, and language together. Daily-Omni 70.1, versus 61.3 for Qwen2.5-Omni-7B. 2.2B parameters.

The decision models d1-2B / d1-Omni-2B output bounding boxes with calibrated probabilities, field extraction, numeric ranges, and multiple-choice probabilities. They average 63.2 across 117 benchmarks, versus 58.8 for d1-3B, with one-third fewer parameters.

Architecturally, the LFM3 language backbone uses a 24-layer hybrid structure: gated short convolutions, sliding-window attention, and global attention. Liquid AI says the architecture was designed with automated hardware-in-the-loop search, optimized for 14 classes of processors, including phones, laptops, Raspberry Pi, cloud CPUs, GPUs, and NPUs. Deployment constraints were baked into the architecture choices, while optimizing both per-parameter capability and on-device inference speed.

LFM3-VL-2B GGUF on LM Studio is about 1.10GB. Developer SourceCodeplz downloaded the Q8 quantization and noted: the file is 1.85GB, with about 1.6GB extra for the 128K context, totaling roughly 3.5GB of memory. He had the model write an HTML page; the result wasn't perfect but was usable, at 97.85 tok/s. The initial GGUF failed to load on LM Studio 0.4.26, which was fixed after updating the runtime. That "the first version always has some issues" detail should be familiar to anyone who works with local models.

The value of these small models is more obvious in specialized scenarios. The company gave a wind turbine inspection example: about 3,000 public photos, SFT+GRPO, a few hours of training on a single A100. Defect classification accuracy went from under 10% to 67%, and detection F1 from 0.003 to 0.11. On the robotics side: navigation R2R 53%, RxR 43%; manipulation VLA averages 96.2% on LIBERO, with action generation under 50ms on Jetson AGX Orin.

LFM3 won't replace frontier models like GPT-6. Its approach is a different one: be good enough at specific tasks while staying small enough to fit on your device. One commenter said their old laptop finally has something to do again. A model that runs in 1GB of memory, can call tools, and handles vision and audio — that trend is more worth noting than the parameter numbers. The entry ticket for local agents may already have dropped to 1GB.