LFM3 shares a small backbone across six models, with a separate route for structured decisions

One trunk branches into six different jobs

LFM3-1.7B generates text. LFM3-VL-2B adds images and video as inputs; LFM3-Audio-2B accepts text or audio and produces text or speech. LFM3-Omni-2B combines text, visual and audio inputs with text or speech output. The branch diagram shows separate training paths from one base, rather than one checkpoint with six interchangeable interfaces. The text card separately describes MOPD distilling six specialist teachers into one student; those teachers are not the six product branches.

The remaining branches, d1-2B and d1-omni-2B, derive from VL and Omni respectively and return structured answers with probabilities. AISubHub’s interpretation is to choose the required input and output contract first: a voice conversation and a bounded document-routing decision pose different engineering problems. “2B-class” also does not mean every model has exactly two billion parameters.

Local processing and global context share 24 layers

LiquidAI reports 13 gated convolutions, with kernel size seven rather than LFM2’s three; five attention layers use a 256-token sliding window, and six attend globally. Head-wise sigmoid gates, RoPE on only the highest-frequency half of each global-attention head, full RoPE in sliding-window layers and input/output normalization complete the described design. Hardware-in-the-loop search covers 14 device/backend combinations; it does not certify every deployment on those devices.

The base was trained on 22T tokens, including a final 3.5T mid-training stage, with a 128K vocabulary. The text card specifies 256K context for input and output together. In the blog’s per-parameter edge comparison with LFM2-1.2B, short-prompt speed matches and 8K-context inference is about 25% faster; it also reports about 1 GB at 4-bit precision. We would budget runtime, context-cache and modality overhead separately, rather than promise a full long-context application fits that footprint.

Text and vision comparisons need their original denominators

For text, the blog reports 35% fewer parameters than LFM2.5-2.6B and WebArena-Verified scores of 38.42 versus 3.02. For vision, LFM3-VL-2B is compared with LFM2.5-VL-3B at 1.9B versus 3.1B parameters and an average across 62 benchmarks rising from 57.6 to 71.4, described as comparable to Gemma 4 12B. These are source-reported comparisons, not an overall ranking or results reproduced here.

The new 164M vision encoder reportedly uses one-third the compute of the previous SigLIP2 encoder. Its separate accuracy comparison with Qwen3.5-2B’s encoder uses the same LFM2.5-230M language backbone and reports about 1/2.2 as many FLOPs. That controlled encoder comparison is not whole-model latency. Combining each sampled frame pair into one temporal slice halves the temporal sequence, without establishing unchanged accuracy for every video.

Speech timing describes components, not the complete conversation

The audio branch adds a causal encoder and LiquidCodec. The blog describes translation across 18 languages and reports UTMOSv2 of 3.83 versus 3.22 for LFM2.5-Audio-1.5B. It states time to first token can be as low as 50 ms, but does not specify that figure’s hardware or complete measurement conditions. First-token timing is not a demonstrated microphone-to-answer delay.

The separate roughly 87-times-real-time figure applies to codec speech decoding on one CPU thread of an Apple M4 Max. It should not be transferred to model generation, network transport or all devices. Omni combines modalities and cites Daily-Omni, WorldSense and AV-SpeakerBench comparisons against larger models. Those named tests do not establish universal multimodal superiority or our own assessment of speech quality.

d1 answers named questions in one forward pass

Instead of asking an autoregressive model to compose a paragraph, d1 accepts a state and multiple questions. The example classifies a refund request, routes a ticket, selects topics and scores urgency. The stated single forward pass concerns answering those questions; it does not remove image/audio preprocessing or make the whole application one operation. d1-omni adds audio input to this decision approach.

LiquidAI reports d1-2B uses one-third fewer parameters than d1-3B and scores higher on the four decision suites presented, including 75.5 versus 64.0 on ImajevBench. “Calibrated” remains a vendor characterization: a probability is not proof that an individual answer is correct. Our deployment recommendation is to validate calibration and routing thresholds on the intended data, including changed distributions and cases needing human review.

Fine-tuning still leaves hardware, use-case and license questions

The notebooks cover SFT for document JSON, DPO for retrieval-grounded answers or refusal, GRPO for a 108-tool phone application, and CPT for Turkish, alongside vision and spoken-command tasks. The blog says every model can fine-tune on one GPU, without specifying GPU memory, sequence length, training duration or cost. A tutorial is not evidence that a resulting application is reliable.

The text card recommends retrieval/search for knowledge-heavy tasks and does not recommend agentic coding or complex multi-step reasoning because of the model’s size. We ran no code or models, audited no notebooks and did not watch the linked robotics video; this article therefore makes no robotics-reliability claim.

For the text checkpoint, the card specifies LFM Open License v1.0. Its text defines a US$10 million annual-revenue threshold and restricts commercial use by entities exceeding that threshold, with a qualified-nonprofit noncommercial/research exception and redistribution obligations. This is not unconditional commercial permission; the exact boundary needs the full terms, and the other five checkpoints’ licenses were not individually checked.