Liquid AI Releases LFM3, a Unified Multimodal Model Family for Text, Images, and Audio

Liquid AI has launched LFM3, a new family of models that processes text, images, and audio within a single shared architecture — a meaningful step beyond the single-modality models the company shipped last year. The launch matters because it gives developers building on-device or edge AI a unified backbone rather than stitching together separate models for each input type.

LFM3 is built on Liquid AI's liquid foundation model (LFM) architecture, which uses structured state-space models rather than the transformer design that underpins most large language models. The company's thesis has always been that non-transformer architectures can deliver competitive performance at lower inference cost, and LFM3 extends that bet into multimodal territory.

What Launched

Liquid AI released a family of LFM3 variants covering three input modalities:

• Text models — the baseline LFM3 series, continuing the company's earlier text-only work
• Audio and omni variants — include native speech generation, Liquid AI's first deployment of voice synthesis
• Vision capabilities — image understanding integrated into the shared model backbone rather than bolted on as a separate encoder

All variants are available today via Ollama, the local model runtime, on Linux, Windows, and Apple silicon (M-series chips). No cloud API has been announced alongside the launch.

The "omni" models combine all three modalities — text, image, and audio — in a single inference path. This mirrors the architecture direction OpenAI took with GPT-4o, though Liquid AI's implementation runs locally rather than through a hosted endpoint.

Why the Architecture Matters

Transformer-based multimodal models typically handle different input types by routing them through separate encoders before merging into a shared latent space. Liquid AI's structured state-space approach processes sequences with linear rather than quadratic computational scaling relative to context length, which can make long audio or high-resolution image inputs cheaper to run.

For developers running models on consumer hardware — laptops, workstations, industrial edge devices — that efficiency difference is real. Running GPT-4o-class audio and vision capabilities locally has not been feasible on most hardware; Liquid AI is betting LFM3 changes that calculus for at least some use cases.

Who This Is For

Liquid AI positions LFM3 for specialized decision-making tasks rather than general conversational use. The practical targets include:

• Manufacturing quality inspection (image + text reasoning)
• On-device voice assistants that cannot route data to the cloud
• Industrial monitoring systems combining sensor audio with visual feeds

The Ollama distribution method reinforces this positioning. Ollama is the primary tool developers use to run open or locally-deployable models — it is not a consumer product. LFM3's launch audience is engineers evaluating whether a non-transformer multimodal backbone can replace multi-model pipelines.

What to Watch

Liquid AI has not disclosed benchmark scores for LFM3 against standard multimodal evaluations such as MMBench or OpenASR leaderboards. Independent benchmarking by the open-source community will arrive quickly given Ollama availability, and those results will determine whether the efficiency claims hold up against transformer-based alternatives at comparable parameter counts. Watch also for whether Liquid AI announces an API endpoint — staying local-only limits the addressable market considerably.