Frequently asked questions.
Yes. The Liquid API serves our LFM and d1 decision models with pay-per-token pricing. Sign up on Liquid Console to get started. The same models are also on OpenRouter and Vercel AI Gateway at the same price. You can try them for free on Playground.
To run models on your own infrastructure, download our open-weight models from Hugging Face. Businesses under $10M in annual revenue can use them in production for free. Above that, you need an enterprise license. For custom deployments or dedicated support, contact our sales team.
Yes. Our open-weight models are released under the LFM Open License, which grants broad, royalty-free rights to use, modify, and distribute them, including commercially. Businesses under $10M in annual revenue can use them in production for free. Above $10M, you need a paid enterprise license, so please contact our sales team.
Our docs cover the full customization and deployment workflow. Liquid Finetune is our open-source repo for data preparation, training (SFT, DPO, GRPO, LoRA, or full fine-tuning), evals, and export to Hugging Face or GGUF. It’s built around Liquid Foundation Models.
For other models, the docs also include guides for widely used open-source tools. These include TRL and Unsloth for fine-tuning, vLLM and SGLang for GPU serving, and llama.cpp, MLX, and Ollama for on-device inference. For help with a custom project, contact our sales team.
Browse every LFM, with benchmarks and specs, on our Models page. Read about the work behind them on our Research page and in News. To start building, see the docs or download the weights from Hugging Face. For updates, join our Discord and follow us on X and LinkedIn.
Yes. Every LFM is free to download and fine-tune, and your fine-tunes stay yours under the LFM Open License. Liquid Finetune is our open-source repo for the full customization loop: data preparation, training (SFT, DPO, GRPO, LoRA, or full fine-tuning), evals, and export to Hugging Face or GGUF. You can also use TRL or Unsloth.
For enterprise projects, our team can customize and deploy models on your hardware, end to end. Contact our sales team to get started.
Yes. You can download our open-weight models from Hugging Face and run them on your own infrastructure, from edge devices to on-prem servers, so your data stays in your environment. Businesses with more than $10M in annual revenue need an enterprise license for production use. Our team can also customize and deploy models on your hardware, end to end. Contact our sales team to learn more.
They are open weight. This means the model weights are publicly available for download, local deployment, and commercial use. This is different from open source, which would include releasing the full training code, data, and methodology. See our pricing for options or get in touch with us for questions.
Liquid models solve for latency, privacy, and cost. They run fast on the hardware you already have, from phones and laptops to a single GPU, so data can stay on the device and costs stay predictable at high volume.
- Speed: our models are built to be the fastest in their class. LFM3-1.7B decodes 1.8 to 2.8 times faster than the next model on every device tested. LFM2.5-24B-A2B is up to 1.7x faster than the next-best model while using the least memory. LFM2.5-DSpark adds up to 3.2x faster inference from H100s to MacBooks. In the data center, a single NVIDIA B300 serves 3.6 billion LFM3 tokens per day.
- Agents and tool use: LFM3-1.7B leads on instruction following and function calling on BFCL and ToolSandbox. On long tasks in browsers and business apps, it outperforms models several times its size. It’s post-trained for agent harnesses like Hermes Agent, Pi, and OpenClaw, and handles email triage, calendar management, web research, and tool search. LFM2.5-24B-A2B decodes 184 tokens/s on an Apple M5 Max. LFM2.5-2.6B holds about 30 tokens/s on a phone, so an agent can run continuously on the device.
- Vision: LFM2.5-VL-3B is especially strong at reading screens and documents, detecting objects, and calling tools on what it sees. It decodes 228 tokens/s on an M5 Max and runs on a phone and in the browser.
- Decisions: our d1 decision models handle classification, routing, scoring, and moderation over text and images. One call returns a calibrated probability for every possible answer, with no text to parse. d1-ultra tops the Artificial Analysis Decision Index and costs $0.042 per million input tokens. The open d1-3B answers in under 50 ms on every device Liquid measured, which is fast enough for live camera input.
- Search: the LFM2.5 retrievers power fast multilingual search across 11 languages over product catalogs, knowledge bases, and support docs, even on a CPU.
- Customization: a short fine-tune on a single GPU adapts a model to your own domain, tools, and workflows.
See them in action in our demos, try d1 live, or explore solutions for industries like automotive, financial services, and healthcare.
Yes, the Liquid API is open to everyone. Sign up on Liquid Console and each new organization gets $2 in free credit. The same models are also on OpenRouter and Vercel AI Gateway at the same price. See the API docs to get started.
To try the models without writing code, chat with them for free on Liquid Playground, or try our decision models in the d1 demos. You can also download every open-weight model from Hugging Face and run it on your own hardware.