# LFM3 Is Free Until You Make $10M a Year. Bootstrappers Win > Liquid's LFM3 is free for commercial use below $10M revenue, with no copyleft and a text model that fits in about 1 GB. What that licence unlocks. - Canonical: https://zeroraised.com/blog/lfm3-10m-revenue-licence - Published: 2026-10-29 - Author: ZeroRaised Team (https://zeroraised.com/blog/authors/zeroraised) - Publisher: ZeroRaised (https://zeroraised.com/) - Section: Growth - Tags: artificial-intelligence, bootstrapping, revenue, developer-tools Liquid AI released the LFM3 family on 27 Oct 2026, and the most interesting part is not the benchmark table. It is the licence. The LFM Open License v1.0 allows free commercial use until your company's annual revenue reaches $10M, with no copyleft and no cap on users ([Liquid AI](https://www.liquid.ai/lfm-license)). For a founder who has been paying per token since the first prototype, that is a different kind of deal. **TL;DR** - LFM3 is open-weight and free to use commercially below $10M annual revenue. - Six 2B-class models cover text, vision, audio, omni and decision work. - The text model fits in about 1 GB at 4-bit and takes a 256K context window. - No copyleft means your fine-tunes can stay proprietary. - Free under a threshold is a distribution strategy, not just a licence. **Disclosure:** ZeroRaised publishes this analysis and is a launch site for bootstrapped products. Facts come from the linked sources as checked on 29 Oct 2026; ZeroRaised has no commercial relationship with the companies named. **Revenue-threshold licensing** is a model licence that grants full commercial rights below a set annual revenue figure and requires a paid commercial licence above it. ## What does the LFM3 licence actually allow? The LFM Open License v1.0 is based on Apache 2.0 ([Liquid AI](https://www.liquid.ai/lfm-license)). Commercial use is free, and those rights end when your company's annual revenue reaches $10,000,000 or more, which the licence calls the Threshold ([Liquid AI](https://www.liquid.ai/lfm-license)). Above it you have to buy a commercial licence from Liquid ([Liquid AI](https://www.liquid.ai/lfm-license)). Three details matter more than the headline number. First, there is no user threshold, so serving a large number of users does not by itself trigger the licence ([Liquid AI](https://www.liquid.ai/lfm-license)). Second, there is no copyleft, so you can keep fine-tuned models proprietary ([Liquid AI](https://www.liquid.ai/lfm-license)). Third, attribution is required: you must retain the notices ([Liquid AI](https://www.liquid.ai/lfm-license)). Research and qualified non-profit use are exempt from the threshold, and the licence terminates automatically if you stop complying ([Liquid AI](https://www.liquid.ai/lfm-license)). A licence that changes at a revenue line is not the same thing as plain Apache 2.0, even when it is based on it. ## What can a 2B model actually do? The LFM3 family covers text, vision, audio, omni and decision models, and the first release sits in the 2B parameter range ([Liquid AI](https://www.liquid.ai/blog/lfm3)). Six 2B-class models were trained from LFM3-1.7B-Base on 22T tokens of pre-training ([Liquid AI](https://www.liquid.ai/blog/lfm3)). Liquid says LFM3-1.7B outperforms LFM2.5-2.6B with 35% fewer parameters, which is a vendor claim rather than an independent result ([Liquid AI](https://www.liquid.ai/blog/lfm3)). The practical targets are bounded and repetitive: classification, extraction, routing between tools and short decisions. That is exactly the work that quietly runs up an API bill in a small SaaS, and exactly the work where a small model can be good enough. Liquid's own example is worth studying. It fine-tuned LFM3-VL-2B to spot defects in wind turbine blades using about 3,000 public photos, a few hours on one A100, and classification accuracy rose to 67% ([Liquid AI](https://www.liquid.ai/blog/lfm3)). Out of the box, the base model and GPT-6 Luna found the defects in fewer than 10% of images ([Liquid AI](https://www.liquid.ai/blog/lfm3)). A 67% model is not a product on its own, but it shows how far a narrow fine-tune can move a tiny model. There is a second data point. MacPaw's Ivan Petrukha reported that across 65 everyday agent tasks on macOS, LFM3-1.7B reached 64% task success against 48% for Qwen3.5-2B, with 58% fewer tokens per task, and got stuck in under 1% of attempts against 15% ([Liquid AI](https://www.liquid.ai/blog/lfm3)). ## Why does zero marginal token cost change product design? The text model fits in about 1 GB at 4-bit precision, has a 256K context window, and every model in the family can be fine-tuned on a single GPU ([Liquid AI](https://www.liquid.ai/blog/lfm3)). It has native support for llama.cpp, MLX and vLLM, and Liquid credits Ollama, LM Studio, Hugging Face and Unsloth in its builder ecosystem ([Liquid AI](https://www.liquid.ai/blog/lfm3)). That changes what you can afford to build. Liquid puts it plainly: "Since tokens are free, dozens of agents can run continuously in the background" ([Liquid AI](https://www.liquid.ai/blog/lfm3)). Think about the features you rejected because the per-token maths did not work: an agent that watches every support ticket, a classifier that re-checks every record every night, a per-user assistant that idles all day. On a hosted API each of those has a meter attached. Locally, the cost is the hardware you already own. The [guide to running a model locally](https://zeroraised.com/blog/how-to-run-an-llm-locally) covers the setup. ## Where is the catch? Two places. The first is scale. At $10M in annual revenue you must buy a commercial licence, so this is a licence for the part of the journey where cash is tight, not a permanent free ride ([Liquid AI](https://www.liquid.ai/lfm-license)). The second is capability. A 2B model is not a frontier model, and vendor benchmarks are not your workload. The [open-weight versus closed model comparison](https://zeroraised.com/blog/open-weight-vs-closed-ai-models) explains where the gap shows up, and [how to choose an LLM API](https://zeroraised.com/blog/how-to-choose-an-llm-api) covers the hosted side of the same decision. Also read the fine print on attribution and termination. A licence that ends automatically on non-compliance rewards the founder who keeps notices intact. ## What is the pricing lesson for your own product? Free below a revenue threshold is good distribution. It lets the smallest customers use the product without a sales call, and it charges only when the customer's business can absorb the cost. Liquid gets adoption from every indie developer who would never sign a contract, and it gets paid by the companies that grow past the line. You can copy the shape without copying the number. A free tier defined by usage or by the customer's stage is easier to defend than a free tier defined by features. The customer who outgrows it already trusts you, which is the cheapest kind of upgrade conversation. The [pricing guide](https://zeroraised.com/blog/how-to-price-a-saas-product) and the [AI pricing guide](https://zeroraised.com/blog/how-to-price-an-ai-product) go deeper on where to put that line. ## What this means for bootstrapped founders - **Pick a bounded task first.** Classification, extraction, routing or a short decision. Do not start with open-ended reasoning. - **Fine-tune on your own examples.** A few thousand labelled items and one GPU can move a small model further than a bigger general model moves on your domain. - **Model the licence into your plan.** Note the $10M threshold, the attribution requirement and the automatic termination. - **Keep a hosted fallback.** The [LLM cost guide](https://zeroraised.com/blog/how-to-reduce-llm-costs) shows how to route only the hard cases to a paid API. - **Price your own product the same way.** Free below a line, paid above it, is a distribution strategy. ## Where ZeroRaised fits ZeroRaised is a launch and discovery site for products built without institutional venture capital or accelerator equity. If you have built something on an open-weight model, or on anything else, you can [submit your product](https://zeroraised.com/submit). The [pricing page](https://zeroraised.com/pricing) explains the current options, and there are no promises of traffic or results. If you want more analysis like this, [join the ZeroRaised newsletter](https://zeroraised.com/newsletter). It is a plain email, not a growth hack. ## Sources - [Liquid AI: Introducing LFM3](https://www.liquid.ai/blog/lfm3) - [Liquid AI: LFM Open License v1.0](https://www.liquid.ai/lfm-license) ## Frequently asked questions ### Is LFM3 really free for commercial use? Yes, below the threshold. The LFM Open License v1.0 grants free commercial use, and those rights end if your company's annual revenue reaches $10,000,000 or more. Above that line you need a commercial licence from Liquid. ### Can I keep a fine-tuned LFM3 model private? Yes. The licence has no copyleft clause, so you can keep fine-tuned models proprietary. You do have to retain the required attribution notices. ### What does the $10M threshold mean for a solo founder? It means a solo founder or small team will almost never reach it. The threshold applies to annual revenue, not to the number of users, and there is no user threshold in the licence. ### What is a 2B model good for? Bounded, high-volume work: classification, extraction, routing between tools and short decisions. Liquid's own example fine-tuned a 2B vision model on about 3,000 photos and raised defect detection accuracy to 67 percent. ### Does the licence cover research and non-profits? Research use and qualified non-profit use are exempt from the revenue threshold. The licence terminates automatically if you do not comply with its terms.