News · 2026-10-28

Liquid releases six LFM3 specialists, including models that decide without chatting

Liquid AI has released six LFM3 specialist checkpoints for local text, vision, speech, and structured decisions, rather than one universal small assistant. The October 27 release’s most distinctive branch returns typed answers and probabilities in one pass, making it a component for automation rather than a chatbot that must write a paragraph for every decision.

Key facts

• The family contains six checkpoints derived from a shared text backbone.
• Liquid describes the 4-bit text weights as about a 1 GB download; full runtime memory is larger.
• The decision models return fields such as choices, extracted values, locations, and scores.
• Detailed primary evidence comes from Liquid’s model card, other official cards, and its license.

Liquid’s launch title, “Every Modality for Specialized Intelligence,” captures the important distinction. Developers choose a specialist, its runtime, and its output interface. The text branch supports agent and tool workflows. Other branches add vision, speech, or combined modalities. The two decision models exchange conversational flexibility for an interface designed to feed ordinary software.

The d1-2B card describes named questions about a supplied state. An application might ask whether an invoice is overdue, extract its total, and locate a relevant object in a picture. The model returns structured results rather than generating each answer word by word. Think of a small inspection station with several labeled gauges, instead of a helpful clerk explaining every observation. That can simplify integration, but software still has to validate the fields and choose when to act.

Probabilities do not settle those choices. Liquid advises setting thresholds using task-specific data. A document extractor that is acceptable for sorting a personal archive may need a much stricter threshold before authorizing a payment. Ground Truth’s calibration lesson explains why a number between zero and one is not automatically a trustworthy confidence estimate. The discriminative-model lesson explains the broader difference between selecting an answer and generating prose.

The text backbone combines short convolutions, local attention, and global attention. Liquid says architecture search considered 14 device-and-backend combinations. That is a hardware-aware design process, not proof of universal compatibility. The official text card reports phone and laptop measurements with a particular quantized format and context length. Its results support a practical local path on those setups, while leaving arbitrary devices and application combinations to be tested.

For storage, the verified company figure is about 1 GB for the 4-bit text weights. The dossier contains no verified file-size total for every other checkpoint, so no size is inferred from parameter counts. A minimum or recommended GPU VRAM requirement is unstated in the reviewed material. The text model also has a CPU execution path, and some tested devices use shared memory. Liquid separately reports 1.6 GB for a full 128K-token context cache: a useful reminder that weight storage is not total working memory.

The maker’s own capability limits matter as much as the footprint. Its text card advises against difficult agentic coding and complex multi-step reasoning. The audio card supports a strong voice-interface story more clearly than across-the-board audio understanding. The vision card describes video sampled at two frames per second, with runtime support still uneven. Two frames per second is not continuous perception of every frame.

Early users provide evidence of interest and friction, rather than independent validation of the benchmark tables. The Hacker News thread includes a specific report of text and Omni variants failing to load in certain local application paths. A Liquid employee suggested another runtime. That exchange does not establish that the original failure was fixed. Other users discuss home automation and personal archives as prospective uses, not completed deployment studies.

A public WebGPU Space is listed as running for browser video captioning, but the dossier did not verify successful inference or the full interface. Its existence should not become a claim that the model works in every browser.

Finally, Liquid’s full license restricts commercial use above $10 million in annual revenue without another agreement. Its exact boundary wording at $10 million is inconsistent, so that case remains unresolved. For a practical first test, the text checkpoint’s documented local path is better grounded than claims about universal browser compatibility. A short task-specific evaluation can reveal errors that a family-wide average hides. The release is a meaningful expansion of small, downloadable automation components. Its value depends on choosing the right task, validating the runtime, and testing the decisions that will actually matter.

Primary source, verified: read the paper →

Ground Truth. (2026, October 28). Liquid releases six LFM3 specialists, including models that decide without chatting. Ground Truth. https://groundtruth.day/news/liquid-lfm3-six-specialists-local-decision-models.html

@misc{groundtruth:liquid-lfm3-six-specialists-local-decision-models,
  title  = {Liquid releases six LFM3 specialists, including models that decide without chatting},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/liquid-lfm3-six-specialists-local-decision-models.html}
}

Topics: open-weights · local-ai · liquid-ai · edge-ai · decision-models · multimodal