Liquid AI releases LFM3, a family of open-weight models built for phones and laptops

What Liquid AI released

Liquid AI released the LFM3 family on 27 October 2026, a set of six open-weight models in the 2-billion-parameter class. The family covers text, vision, audio, an omni model and a decision model, all trained from one 1.7B-parameter base.

The models are available on Hugging Face under the LFM Open License. The headline text model, LFM3-1.7B, is described by Liquid as a text model for local agents and specialized applications, based on its new tri-hybrid architecture.

The specifications

LFM3-1.7B has 1.73B parameters across 24 layers, with a mix of 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers. It has a 256K-token context window and a 128,000-token vocabulary, covering English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian and Polish.

Liquid says at 4-bit precision the model takes about 1 GB of memory, so it fits on any phone or laptop. The design mixes short convolutions with sliding-window and global attention, tuned on devices from phones and a Raspberry Pi 5 to laptops. Liquid says LFM3-1.7B beats its earlier LFM2.5-2.6B with 35% fewer parameters.

Why on-device models matter

Small models that run locally let agents work without sending data to the cloud, which matters for privacy, cost and offline use. Liquid's release is part of a broader push toward on-device and edge AI, alongside the browser-based local AI standard Opera announced the same week.

The company says LFM3-1.7B is post-trained for agentic harnesses such as Hermes Agent, Pi and OpenClaw, positioning it for tool calling, structured outputs and domain-specific fine-tuning. Liquid cautions that the model is not recommended for agentic coding or complex multi-step reasoning due to its small size.

Early-adopter feedback will matter more than benchmark claims: developers testing the previews will quickly find what the models and APIs can and cannot do, and that feedback will shape both products.

What to watch

Watch for independent benchmarks of the LFM3 family, since the performance figures Liquid published are its own and have not been independently verified.

Watch whether LFM3 gains adoption in local-agent and edge deployments, which would show whether the 2B class is becoming a useful tier for on-device AI rather than a curiosity.

Watch how quickly competitors respond in the same tier or standard, and whether the early previews translate into real developer adoption over the next quarter.

The decisive test will be whether real developers ship real products on these foundations and whether the economics of local AI hold up against cloud inference at scale.

The agentic angle

Liquid says LFM3-1.7B is post-trained for agentic harnesses such as Hermes Agent, Pi and OpenClaw, with a 256K-token context window. That positions it as a model an agent can run locally and keep in context across long tool-use sessions, rather than as a front-end for a bigger cloud model.

The choice to name concrete agent harnesses is a sign of where small-model demand is heading: developers want models they can hand to an agent loop on a laptop, without paying per-token cloud costs or sending prompts off-device. Liquid also published guidance on fine-tuning for tool calling, structured outputs and domain-specific tasks.

The market context

LFM3 lands in a crowded small-model market that includes Google's Gemma family, Microsoft's Phi and Qwen's small checkpoints. Liquid's differentiator is its Liquid Neural Network heritage and the tri-hybrid architecture mixing convolutions and attention, which it says delivers LFM2.5-level performance at 35% fewer parameters.

The release also fits a broader on-device push: Opera announced WebAI, a browser standard for local models, the same week, and Apple has been shipping on-device foundation models. If small open-weight models reach a quality floor where local agents are genuinely useful, Liquid is well positioned in that tier.

The tri-hybrid architecture

The tri-hybrid design is LFM3's signature. Most small models use one attention mechanism; Liquid mixes three: gated short convolutions that handle local patterns efficiently, sliding-window attention that keeps a rolling context, and global attention that sees the whole sequence. That division of labor aims to get long-context capability without the quadratic cost of full attention at every layer.

For developers running on phones or laptops, the practical consequence is a model that sustains a 256K-token context in about a gigabyte of memory. Liquid publishes quantization variants for GPU, CPU and Apple Silicon, plus ONNX for cross-platform inference, so the same checkpoint can run in many runtime environments.

The caveats

Liquid's performance claims are its own: the company says LFM3-1.7B beats LFM2.5-2.6B with 35% fewer parameters, but those results have not been independently reproduced. The model card also warns against using the small model for agentic coding or complex multi-step reasoning.

Small models are also only as good as their training data and fine-tuning. Liquid recommends retrieval or search for knowledge-heavy tasks rather than relying on the model's parametric memory, which is a standard caveat for the 2B class.