What happened
Liquid AI announced LFM3 on October 27, 2026 as a new generation of Liquid Foundation Models. The company positions it as a starting point for specialized intelligence: rather than one large general-purpose model, a builder selects a model for the required input and output modality, adapts it to local data, and deploys it where the application runs. Base and post-trained weights are open through Hugging Face, while the family is also available through the Liquid AI API and Playground.
The release begins with six models in the approximately 2B parameter range, all branching from LFM3-1.7B-Base. LFM3-1.7B is a text-input, text-output model with a 256K-token context window. LFM3-VL-2B accepts text, images, and video and returns text. LFM3-Audio-2B takes text or audio and can return text, speech, or both. LFM3-Omni-2B combines text, image, video, and audio input with text and speech output. The d1-2B and d1-omni-2B variants are decision models: respectively based on the vision and omni models, they return calibrated answers rather than generated text.
Underneath is a tri-hybrid architecture combining gated short convolutions, sliding-window attention, and global attention. Liquid AI says LFM3-1.7B-Base has 24 layers: 13 gated short-convolution layers, five sliding-window attention layers with a 256-token window, and six global-attention layers. It was pre-trained on 22T tokens. The final 3.5T-token mid-training stage increased the share of agentic traces and tool calls and extended context to 256K tokens. The company reports that, with 8K tokens already in context, LFM3-1.7B is about 25% faster than LFM2-1.2B per parameter across edge devices.
The text model is post-trained for agent harnesses including Hermes Agent, Pi, and OpenClaw, and Liquid AI recommends it for local agents, tool calling, structured outputs, and domain-specific fine-tuning. The company says it outperforms LFM2.5-2.6B using 35% fewer parameters, fits in about 1 GB at 4-bit precision, and scores 38.42 on WebArena-Verified versus 3.02 for that predecessor. These are vendor-reported comparisons. The vision model adds video input; the audio model adds speech input and output, tool calling, and support across 18 languages. Liquid AI says every model can be fine-tuned on a single GPU. It lists day-one support in transformers, llama.cpp, MLX, and vLLM, alongside partners including OpenRouter and Vercel.
What the archive adds
This is a new family-level release on a known Liquid AI thread. Our earlier note on d1-ultra examined a multimodal decision interface that returns calibrated probabilities. Our note on LFM2.5-24B-A2B focused on an earlier open model artifact and launch path. LFM3 adds a common 2B-class backbone and six modality-specific branches.
Trend signal
LFM3 supports the pattern that harness architecture outlives model choice. Liquid AI explicitly targets existing agent harnesses and multiple inference engines rather than tying the release to a proprietary execution environment. That does not prove interchangeability in production, but it gives teams an alternative model family to evaluate without redesigning the surrounding workflow.
Connections to other events
The release follows Liquid AI's earlier d1-ultra decision-model work, but broadens that product idea into d1-2B and d1-omni-2B alongside generative text, vision, audio, and omni models. It also extends the company's previously covered open-weight LFM2.5 line from a larger mixture-of-experts release toward smaller models intended for edge and local deployment.
Why it matters
For builders, LFM3 creates a clearer evaluation unit than a generic “multimodal model.” A product team can test the text model for tool invocation or structured output, the vision model for image or video understanding, the audio model for voice interaction, and the decision variants where a calibrated response is more useful than prose. The shared backbone may reduce the operational cost of experimenting across those modalities, although the source does not quantify that cost.
The local angle is consequential as well. A reported 1 GB 4-bit footprint for the text model, plus named support for common local and server inference stacks, makes it plausible to test an agent on a laptop or constrained deployment target before committing to hosted inference. The important decision remains empirical: measure task completion, latency, memory use, and failure behavior on the actual workflow.
Limits and uncertainty
All performance figures in the announcement are Liquid AI's own claims. The source gives selected comparisons, including WebArena-Verified and vision benchmarks, but does not provide an independent evaluation, total deployment cost, rate limits for the API, license terms, or a complete comparison with other open models on a team's specific workload. “Fits in about 1 GB” refers to the text model at 4-bit precision and should not be generalized to every LFM3 variant or runtime. Compatibility with an inference engine also does not establish feature parity, throughput, or production reliability.
Skill to build now
Build a small modality-specific evaluation instead of treating LFM3 as one model. Pick a real task that has a bounded success criterion: a structured tool call, a document-image extraction, a short audio-to-action flow, or a calibrated routing decision. Run the relevant LFM3 branch through the inference stack you already operate, log output validity and latency, and compare it with the model currently behind the same harness. Keep the prompt, tools, validation rules, and test set fixed.
Reader takeaway
LFM3 is an open-weight attempt to make small local AI systems practical across text, vision, audio, combined multimodal interaction, and decision outputs. Its value for a team will depend less on the shared architecture than on whether the specific branch improves a measurable task inside an existing harness.
Open question
Can the agent-oriented gains Liquid AI reports for LFM3-1.7B survive independent, production-like evaluations once tool reliability, long contexts, and local hardware constraints are included?