LFM3 models:
small models, specialized intelligence.

THE DIRECT ANSWER

What is LFM3 by Liquid AI?

LFM3 is a family of lightweight foundation models built on Liquid AI's new 1.73-billion-parameter, 24-layer tri-hybrid backbone. The October 27, 2026 launch included six post-trained variants for language, vision, audio, omni experiences and decision-making, plus a separate base checkpoint for adaptation. Rather than requiring one large general-purpose network to handle every job, the family offers a small checkpoint matched to each input/output contract.

All variants trace back to LFM3-1.7B-Base. Specialized vision and audio branches add their own encoders and components. The decision models are derived from the vision or omni branches and answer questions through calibrated, typed outputs without generating a chat response token by token. This distinction is essential when comparing models, context limits, memory and accuracy.

Liquid AI says the 1.7B base was trained on 22 trillion tokens, including a curated final 3.5T-token mid-training stage emphasizing agentic traces and tool calls. These are developer-reported training details, not independently audited data. Read the technical release notes ↗

ORIGINAL SXF INTERACTIVE EXPERIENCE

Which LFM3 model should you actually use?

Select the real task—not just the model name. This SXF Model-Fit Explorer chooses a checkpoint based on documented input/output capabilities, shows practical acceptance checks, and links to the official model card. It runs entirely in your browser, without a model inference call, upload or remote API.

No recommendation replaces a test using your own inputs. The full preprocessing pipeline, quantization, tool access, safety constraints and actual device memory must be evaluated separately.

SIX RELEASED VARIANTS

LFM3 model family comparison: inputs, outputs and use cases

These are the six specialized launch models. The base checkpoint is separate and is intended for further training, not counted as a seventh post-trained assistant. Context lengths in this table are individual model-card claims where available; “see model card” means no family-wide limit is inferred.

WHAT CHANGED UNDER THE HOOD

LFM3 tri-hybrid architecture: 13 + 5 + 6 layers

Liquid AI's 24-layer LFM3 language backbone mixes 13 gated short convolutions, 5 sliding-window attention layers (256-token local window) and 6 global attention layers. These are different processing operators within a shared backbone, not three separate full-size models.

Other changes include head-wise gated attention, partial RoPE on global-attention heads and sandwich normalization. The vocabulary has 128,000 tokens. Liquid AI used a hardware-aware search across 14 device/backend configurations; the company reports LFM3-1.7B is about 25% faster than the comparison baseline per parameter at 8K context. That is not a universal 25% application speed gain. Architecture evidence ↗

PUBLISHED NUMBERS, NOT SXF MEASUREMENTS

LFM3 benchmarks: where are the improvements?

Below are selected Liquid AI-published results. Compare numbers only within the same named test and the reported setup; differences across model families, prompt policies or tools do not establish an overall winner. Original detailed benchmark tables ↗

For vision comparison, the published LFM3-VL-2B overall average (71.4) is near the vendor's Gemma 4 12B result (71.7), but benchmark composition and inference settings matter. These are not direct independent tests performed by SXF. For production evaluation, track accuracy by your actual task—receipts, diagrams, screenshots, spoken commands—not only a single composite average.

WEIGHTS ARE NOT TOTAL MEMORY

Can LFM3 run locally? RAM, VRAM and quantization

Liquid AI says the LFM3-1.7B text model fits in about 1 GB with 4-bit weights. That is the model-weights footprint, not a minimum system-RAM requirement. Token buffers, context length, KV cache, execution kernels and runtime overhead add memory. Vision and audio checkpoints also need encoders, decoders and preprocessing.

Runtime choices: the official text model card lists BF16 with vLLM, SGLang or Transformers; quantized GGUF for llama.cpp/Ollama/LM Studio; and MLX for Apple silicon. Availability and instructions differ by variant. Do not assume the same quantization support, throughput, context length or speed across all six checkpoints.

Before buying hardware, measure the exact checkpoint plus tokenizer, vision/audio preprocessing, prompt size, batch size and peak memory. A Raspberry Pi or phone appearing in the developer's architecture-search devices is not a guarantee that every multimodal variant works comfortably on every such device.

SOURCE-ALIGNED QUICK START

How to run LFM3 locally with Python

The official LFM3 release includes starter code for the text model. In a Python environment with the required dependencies and compatible hardware, load the model from Hugging Face. Use version-pinned environments for production; models using trust_remote_code=True execute repository-provided Python code, so review and pin the model revision before using it with private data.

python -m pip install -U transformers accelerate

from transformers import pipeline

generator = pipeline(
    "text-generation",
    model="LiquidAI/LFM3-1.7B",
    trust_remote_code=True,
)
messages = [
    {"role": "user", "content": "Return JSON with a title and one-sentence summary of: local AI models."}
]
result = generator(
    messages,
    do_sample=True,
    temperature=0.1,
    top_k=50,
    repetition_penalty=1.1,
    max_new_tokens=512,
)
print(result[0]["generated_text"][-1]["content"])

For images and video, load LiquidAI/LFM3-VL-2B with its official processor; Liquid AI documents transformers>=5.21.0 and vision dependencies. For speech, use the audio model's processor and LiquidCodec (documented transformers>=5.20.2). For combined inputs, follow the omni model card; swapping the text checkpoint for a multimodal one in this simple pipeline is insufficient.

Official links: text · vision · audio · omni · decision.

BUILD SPECIALIST MODELS

How to fine-tune LFM3 for your own domain

Liquid AI provides examples of single-GPU fine-tuning across the family. Choose the post-trained variant whose interface fits the task, or start from LFM3-1.7B-Base if your project needs its own training objective. The published notebooks include supervised fine-tuning (SFT), direct preference optimization (DPO), group relative policy optimization (GRPO) and continued pretraining (CPT)—not every technique is equally appropriate for every task.

SXF evaluation protocol: (1) freeze a held-out test set; (2) save model revision, runtime and preprocessing settings; (3) report a baseline before tuning; (4) fine-tune on non-overlapping examples; (5) compare performance and cost against the same acceptance criteria; (6) evaluate failure cases, safety and calibration. A specialized model that improves one score but fails under distribution shift is not yet deployment-ready.

Vendor example: Liquid AI adapted LFM3-VL-2B to inspect wind-turbine blade images using around 3,000 training photos and tested on around 1,400 held-out images. The company reports classification accuracy reaching 67% after tuning, but this result is tied to that dataset and procedure—not a generic improvement guarantee. Methodology and project details ↗

PREVIOUS GENERATION

LFM3 vs LFM2.5: what actually improved?

The headline claim that LFM3-1.7B outperforms LFM2.5-2.6B with roughly 35% fewer parameters comes from Liquid AI's evaluations. It does not mean every new branch beats every previous checkpoint on every task. If your goal is local vision latency rather than image comprehension, see SXF's separate LFM2.5-VL-DSpark speculative decoding guide; DSpark is a drafter for LFM2.5-VL-3B, not a member of the LFM3 family.

CHECK BEFORE DEPLOYMENT

LFM3 limitations, safety and open-weight licensing

• Small model limits: LFM3-1.7B is suited to narrow agents but the official model card cautions against complex multi-step reasoning and agentic coding. Use retrieval for knowledge-intensive tasks.
• Modality-specific contexts: vision and speech branches do not inherit the text variant's 256K context window. Verify each model card and engine.
• Vendor measurements: LFM3 accuracy, speed, fine-tuning and token counts here reflect Liquid AI's reported setups. SXF has not run an independent LFM3 benchmark.
• No universal hardware promise: a 1 GB quantized weight claim is not proof that an entire agent, image pipeline or 256K KV cache fits in 1 GB of device RAM.
• Licensing: Hugging Face cards label the checkpoints lfm1.0 / LFM Open License v1.0. Open weights are not necessarily an unrestricted OSI-approved open-source license; review conditions before redistribution or commercial use.
• Privacy and security: on-device inference can keep raw data on a device, but downloaded remote code, tool permissions, logs and external inference calls still require review.
• Decision calibration: a probability score is not automatically a correct probability after deployment or distribution shift. Test on target population data and set thresholds against error costs.

SEARCH QUESTIONS ANSWERED

Frequently asked questions about Liquid AI LFM3

PRIMARY-SOURCE REGISTER

Official LFM3 documentation and methodology

Research reviewed November 1, 2026; original announcement October 27, 2026. The source documents the capabilities, model cards, benchmark configurations, checkpoint restrictions and vendor-reported metrics. The model-fit explorer, evaluation protocol and interpretation are original SXF editorial contributions, not new experimental measurements.

Continue researching: Open-weight AI models · Run models locally · LLM quantization · LFM2.5-VL-DSpark · Original SXF launch signal.