LFM3 Pricing

You can download and run LFM3 for $0 in license fees if your organization has less than $10 million in annual revenue. Liquid AI had not published a per-token application programming interface (API) price for LFM3 on its pricing page or on OpenRouter on October 27, 2026.

Liquid AI launched LFM3 on October 27, 2026, releasing model weights across several repositories including LFM3-1.7B. At $10 million or more in annual revenue, commercial use needs a paid license from Liquid AI.

Liquid AI publishes per-token rates for other models such as LFM2.5-24B-A2B but none for LFM3, so Liquid AI's documentation gives no way to calculate a hosted LFM3 bill. Self-hosting cost comes down to memory, sized in the table below.

What LFM3 Costs

LFM3 software licensing costs nothing for organizations with annual revenue below $10 million, while hosted API pricing remains unannounced. The table lists the four cost paths on Liquid AI's pricing page.

Under the self-hosted open-weights tier, developers can download, fine-tune, and deploy LFM3 models across edge devices, mobile operating systems, or on-premise hardware without paying royalties. Proprietary fine-tunes remain private, and the license carries no copyleft provisions. Teams using this path incur only their own hardware and compute expenses, which the open weights models cost guide covers.

For hosted API access, Liquid AI's pricing page lists three models (d1, d1-ultra and LFM2.5-24B-A2B) but contains no row for LFM3. Liquid AI gives every new organization $2 in free credit through Liquid Console. Liquid AI's pricing page does not say whether the $2 credit applies to LFM3.

The License Threshold Under the LFM Open License v1.0

The free commercial grant in the LFM Open License v1.0 stops at an annual revenue threshold. The license agreement defines the financial threshold as annual revenue of 10 million United States dollars ($10,000,000) or more. Any commercial use by a qualifying entity that meets or exceeds this number is explicitly excluded from the grant of rights.

The license defines Commercial Use as any use for direct or indirect commercial advantage or monetary compensation. Furthermore, the agreement applies the threshold to the entire Legal Entity, which encompasses parent corporations, subsidiaries, and sister companies under common control. Control is defined as holding 50 percent or more of the outstanding shares or the operational power to direct company management. The license counts every entity under common control, so a small subsidiary of a group at or above $10 million is treated as meeting the threshold.

Non-commercial research, academic study, and qualified non-profit organizations remain exempt from the revenue cap. Qualified 501(c)(3) entities and equivalent educational or scientific organizations can use LFM3 royalty-free for non-commercial or research purposes regardless of revenue.

The pricing page says free for 'less than $10M', while the license sets the Threshold at '$10,000,000 or more'. A company at exactly $10,000,000 in annual revenue falls under the license restriction and should ask Liquid AI for written clarification. This summary represents a plain-language reading of the documentation rather than legal advice.

Published Liquid API Rates for Other Models

The rates below are Liquid AI's published prices for other models. None of them is an LFM3 price. Liquid AI notes that per-token rates remain identical whether an organization queries models via Liquid Console, OpenRouter, or the Vercel AI Gateway.

On OpenRouter's model list on October 27, 2026, Liquid AI models were LFM2.5-24B-A2B, LFM2.5-2.6B and a free LFM2.5-2.6B variant with a 65,536-token context window. No LFM3 model was listed.

A workload of 1 million uncached input tokens and 1 million output tokens costs $0.02 on d1, $0.042 on d1-ultra, $0.75 on LFM2.5-24B-A2B and $0.85 on LFM2.5-2.6B. None of that tells you what LFM3 will cost.

What Self-Hosting Costs in Memory

Self-hosting adds no license fee below $10 million in revenue. Memory decides which devices can run LFM3-1.7B. Layer3Labs' sizing calculation, shown in the table below, adds up the random-access memory (RAM) for weights at three precisions plus a full 128K cache.

The theoretical memory needed for model parameters uses the standard sizing formula: weights memory in gigabytes (GB) = parameters × bits per weight / 8 / 1,000,000,000. Applying this formula to LFM3-1.7B's 1.73 billion parameters yields 0.86 GB at 4-bit quantization, 1.73 GB at 8-bit precision, and 3.46 GB at 16-bit Brain Floating Point (BF16) precision.

Liquid AI's LFM3 launch announcement says a full 128K-token context needs 1.6 GB of Key-Value (KV) cache in llama.cpp at Q4_0 quantization. Add that to the weights and a 4-bit setup needs about 2.46 GB.

A standard edge device or laptop with 4 GB of available system memory has room on paper for the 4-bit and 8-bit configurations with a full 128K context. However, that same 4 GB device fails to host the BF16 precision model with a 128K context, as the combined 5.06 GB footprint exceeds capacity. These figures represent an arithmetic planning baseline covering weights and KV cache, excluding runtime overhead, operating system processes, and dynamic memory activations.

Limits of the Free Path

Three limits end the free path.

• Annual revenue of $10 million or more: the free grant does not cover commercial use, so contact Liquid AI for a paid license priced around your deployment scale.
• Low-latency, high-volume serving: the Liquid AI launch blog points to Liquid RIPL (Real-time Inference Pipeline) for these workloads and lists no price.
• Agentic coding or complex multi-step reasoning: the LFM3-1.7B documentation does not recommend the model for either.

What This Changes for Your Budget

At Layer3Labs, we build custom agents and workflow automation for business operations. In our internal sizing analysis, we calculated the memory LFM3-1.7B needs using the sizing formula shown in the memory table. Our arithmetic puts LFM3-1.7B at about 2.46 GB with weights at 4-bit and a full 128K cache, or 3.33 GB at 8-bit, before runtime overhead. The license adds no fee below $10 million in annual revenue.

Plan LFM3 as a self-hosted model with no license fee if your organization has less than $10 million in annual revenue. Check that your hardware has the memory shown in the table. Treat the LFM3 API per-token rate as entirely unpublished. Before budgeting hosted inference, check Liquid Console for current LFM3 endpoint pricing.

Who should not rely on this: organizations at $10 million or more in annual revenue. If you need a hosted price today, use a model with a published rate from our AI model pricing index.

What would change our recommendation: Liquid AI publishing a definitive per-token API price card for LFM3 on its pricing page, or altering the revenue thresholds inside future revisions of the LFM Open License. Compare your annual revenue with the $10,000,000 Threshold in the LFM Open License v1.0, then check your hardware against the memory table above. The LFM3 explained guide covers which model to download.