Liquid AI LFM3: six 2B-class open-weight models for text, vision, audio, omni and decisions, free only below $10 million in revenue

Liquid AI LFM3 is the company's new generation of small models, released on October 27, 2026 under the banner "Every Modality for Specialized Intelligence." Instead of one model, LFM3 is a family: a 1.7-billion-parameter base model and six post-trained versions of it for text, vision, audio, a combined "omni" model, and two "decision" models that answer questions with probabilities instead of generating text. All are open-weight on Hugging Face. As with Liquid's LFM2.5-24B-A2B, released on October 13, 2026, the license is free for small companies and not for large ones.

This story is based on Liquid AI's October 27 launch post, its Hugging Face blog post, the Hugging Face model cards for LFM3-1.7B, LFM3-VL-2B, LFM3-Audio-2B, LFM3-Omni-2B, d1-2B and d1-Omni-2B, Liquid's API pricing page and the LFM Open License v1.0 text on Liquid's site, checked October 28 and 29, 2026.

The short version

• What "every modality" means: one shared backbone, with separate models for text, vision (including video), audio (speech in and out), omni (all of these together) and decisions.
• Sizes: 1.7B to 2.2B parameters each, per the model cards and Liquid's tables.
• Context: 256K tokens for the text model, 128K for vision and the two decision models, 32K for audio and omni.
• Where it runs: phones, laptops, a Raspberry Pi 5 and servers, according to Liquid; builds for llama.cpp, Ollama, LM Studio, MLX, ONNX Runtime, vLLM and SGLang are on Hugging Face.
• License: LFM Open License v1.0. Not licensed for commercial use by companies with $10 million or more in annual revenue.
• Performance: every speed and benchmark figure comes from Liquid and hasn't been reproduced independently.

What does "every modality" mean in LFM3?

It means one base model with a version for each kind of input and output, not one model that does everything. Liquid's post says LFM3 brings "text, vision, audio, omni, and decision models onto one new shared backbone," with six 2B-class models trained from LFM3-1.7B-Base. Liquid also publishes the base model itself for custom post-training.

The decision models, d1-2B and d1-Omni-2B, are the unusual ones. According to Liquid's Hugging Face post, you give a d1 model a situation and a list of questions, and it answers them all "in a single forward pass, each with a calibrated probability": yes or no, a choice, a score, a pointer to an object, a numeric range. The d1-2B card describes the output as "typed answers with their probabilities, with no generated tokens." Liquid's examples include a robotics setup, where a structured answer is more useful than prose.

The model cards give the sizes: 1.73B parameters for the text model, 1.92B for the vision model and for d1-2B, 2B for the audio model and 2.17B for the omni model and for d1-Omni-2B. Both decision-model cards give a 128K-token context.

Last verified: October 29, 2026

How is LFM3 built?

Liquid calls the design "tri-hybrid." Its 24 layers combine 13 gated short-convolution layers, 5 sliding-window attention layers with a 256-token window, and 6 global attention layers, according to the launch post. Most layers only look at nearby tokens, which keeps memory low as a conversation grows. Liquid says a 128K-token context needs 1.6 GB of cache, against 4.3 GB for Qwen3.5-4B in its comparison.

The base model was pre-trained on 22 trillion tokens, with the last 3.5 trillion used to stretch the context window to 256K, followed by 1.75 trillion tokens of supervised fine-tuning, Liquid says. The vision and audio models use new in-house encoders: a 164M-parameter vision encoder, and a fully causal audio encoder that Liquid says can start answering in "as low as 50 ms."

What is LFM3's license, and is it free for commercial use?

It's free for commercial use only if your organization makes less than $10 million a year. Every LFM3 card checked lists the LFM Open License v1.0. That license defines the "Threshold" as "annual revenue of 10 million United States dollars ($10,000,000) or more," and says any commercial use of the model or a derivative "by a Legal Entity that exceeds the Threshold is not licensed." Larger companies are directed to Liquid's sales team for a separate commercial license.

The rest follows Apache 2.0's pattern:

• Below $10 million: use, modify, fine-tune and redistribute, keeping the license and notices.
• Non-profits: qualified 501(c)(3) organizations, or foreign equivalents, are exempt from the threshold for non-commercial or research use.
• Fine-tunes: you can license your own changes differently, but the model as a whole stays under the LFM license, so the revenue limit follows it. Liquid's FAQ says there's no copyleft requirement, so a fine-tune can stay proprietary.
• Hosting platforms: the FAQ says a platform redistributing the model "does not modify or waive any license terms for end users."
• No acceptable-use policy: the license has no list of banned uses; its only restrictions are the revenue threshold and trademark rules.

That's the same license as LFM2.5, so nothing has changed for Liquid's customers. It also means LFM3 isn't open source under the Open Source Initiative's terms, a distinction our open-weight vs open-source explainer covers. For how the $10 million annual limit compares with Mistral's, Qwen's and Meta's terms, see our open-weight license comparison.

Where does LFM3 run?

Small enough for a phone, by Liquid's account. Liquid says the text model "fits in about 1 GB at 4-bit precision," and the LFM3-1.7B card says that makes it fit "on any phone or laptop." Liquid's post lists 14 device and backend combinations it ran, including a Samsung Galaxy S26 Ultra, a OnePlus 15, a Raspberry Pi 5, an Apple M4 Max, AMD Ryzen AI laptops, cloud CPUs, GPUs and Qualcomm's Hexagon NPU.

The Hugging Face cards list these builds:

• LFM3-1.7B: GGUF for llama.cpp, Ollama and LM Studio; 4-bit MLX for Apple Silicon; ONNX; FP8 and NVFP4 for SGLang; BF16 for vLLM, SGLang and Transformers; and Docker Model Runner.
• LFM3-VL-2B and LFM3-Audio-2B: GGUF, 4-bit MLX and ONNX, plus BF16 for vLLM, SGLang and Transformers. The vision card says video is sampled at 2 frames per second.
• LFM3-Omni-2B: BF16 for SGLang (text output) and Transformers, and GGUF for llama.cpp, including a real-time speech server.
• d1-2B: Transformers with Liquid's custom code, vLLM, SGLang and quantized builds for llama.cpp, Ollama and LM Studio.
• d1-Omni-2B: BF16 weights for Transformers with Liquid's custom code.

Liquid's launch post says the models are also available on its Playground and the Liquid AI API. Its pricing page lists only LFM3-1.7B among them, at $0.08 per million input tokens, $0.02 for cached input and $0.20 for output, on Liquid Console only.

If you haven't run a local model, our Ollama walkthrough covers the basics, and our laptop memory guide explains how much RAM a model of this size needs.

How fast and how good is LFM3? (vendor-reported)

All figures here were published by Liquid AI on October 27, 2026 and haven't been independently reproduced.

• On-device speed: LFM3-1.7B decodes "1.8 to 2.8 times faster than the next one" on three devices, measured with llama.cpp at Q4_0 with a 32K-token context. The post doesn't spell out which models it was compared with in that passage.
• Server throughput: 3.6 billion tokens a day on one Nvidia B300 at 1,024 concurrent requests (SGLang, BF16).
• Text quality: Liquid compares LFM3-1.7B with LFM2.5-2.6B, MiniCPM5 2B, Gemma 4 E4B, Qwen3.5 4B and Qwen3.5 9B. It says LFM3-1.7B "beats Qwen3.5-9B on long tasks in browsers and business apps (WebArena-Verified, AutomationBench)" and outperforms LFM2.5-2.6B with 35% fewer parameters. It scores 38.42 on WebArena-Verified, against 3.02 for its predecessor. On long-context recall (RULER at 128K), it scores 65.56 to Qwen3.5-4B's 84.63, by Liquid's own table.
• Vision: an average of 71.4 across 62 vision evaluations for LFM3-VL-2B.
• Decisions: d1-2B scores 75.5 on ImajevBench against 64.0 for Liquid's earlier d1-3B, while using a third fewer parameters, Liquid says.

The numbers suggest strength on agent and tool tasks for the size, with weaker long-context recall than larger rivals. Treat them as Liquid's claims until outside evaluations appear.

What's confirmed, reported and uncertain

• Confirmed (Liquid AI, Hugging Face): the October 27 release; the six post-trained models and the base model; parameter counts, inputs, outputs and context on the cards; the architecture; the LFM Open License v1.0 on each card checked and its $10 million annual-revenue threshold; the listed runtimes and builds; LFM3-1.7B's API price on Liquid's pricing page.
• Vendor-reported: all speed, memory and benchmark figures, and the device list.
• Uncertain: API prices for the other LFM3 models, which Liquid's pricing page doesn't list; how Liquid prices commercial licenses above the threshold; independent quality results.