# LFM3: Models, Context Window, API Access and Licensing

Liquid AI's small multimodal models make local specialization more practical, but memory, licensing and deployment choices still matter.

By Paras Madan | 2026-10-28


Liquid AI released LFM3 on October 27, bringing text, vision, audio and decision models into a compact, shared model family. Open weights and API access are available, while the text model supports a 256K-token context window. The practical question is which jobs can now stay on your own hardware.

## What LFM3 includes

The [LFM3 announcement](https://www.liquid.ai/blog/lfm3) describes six 2B-class models built around the same foundation. Developers can choose a modality-specific model rather than loading every capability into every application. Liquid says the models support single-GPU fine-tuning, which makes customization part of the release's central proposition.

This matters for repetitive tasks with a clear definition of success. A support-ticket classifier needs reliable labels and escalation rules. It does not necessarily need the breadth of a large general-purpose assistant, provided it can recognize requests outside its scope.

## LFM3-1.7B specifications

The [official text-model card](https://huggingface.co/LiquidAI/LFM3-1.7B) gives the clearest starting point for a deployment decision.

| Property | LFM3-1.7B |
| --- | --- |
| Parameters | 1.73 billion |
| Context window | 256K tokens, input and output combined |
| Architecture | 13 gated short-convolution layers, 5 sliding-window attention layers and 6 global-attention layers |
| Language coverage | 16 languages, including English, Hindi, Chinese and Arabic |
| License | LFM Open License v1.0 |

The mix of short convolutions, local attention and global attention gives different parts of the network different ways to process a sequence. For builders, the useful consequence to test is how inference behaves as the prompt gets longer, rather than treating parameter count as a complete measure of cost.

Liquid reports that the text model takes about 1 GB at 4-bit precision. **That number is not a complete application's memory budget.** Runtime allocations, conversation state and tool processes still need space. A 256K context limit should not be read as a promise that every phone can comfortably run the maximum-length workload.

## What the benchmarks establish

Liquid's published WebArena-Verified result is 38.42 for LFM3-1.7B, compared with 3.02 for LFM2.5-2.6B and 28.94 for Qwen3.5-9B. These are vendor-reported results for a particular evaluation, not an independent guarantee that the smaller model will complete your browser workflow more reliably.

The model card does not recommend LFM3-1.7B for agentic coding or complex multi-step reasoning, and suggests retrieval or search for knowledge-heavy tasks. That is a useful limit: a strong result on tool use does not make the model a universal replacement for a larger coding assistant.

For a ticket-routing application, I would evaluate incorrect escalations and missed urgent tickets before comparing aggregate benchmark scores. The economically useful measure is the cost of a correctly handled ticket, including any human review it creates.

## API access and commercial licensing

Liquid lists LFM3 as available through its API and Playground, alongside downloadable weights. Its public [pricing page](https://www.liquid.ai/pricing) did not expose an LFM3-specific rate card in the version available for this review. Check the current model entry before budgeting; older LFM2.5 prices should not be applied to LFM3.

The published [license](https://www.liquid.ai/lfm-license) also places a commercial condition on open-weight use. Businesses exceeding $10 million in annual revenue need a separate commercial license. The license permits private modifications and has specific research and nonprofit provisions. Open weights therefore do not mean unrestricted Apache 2.0 licensing.

Local inference can remove a hosted per-token bill, but hardware, maintenance and any applicable license remain costs. And a locally running agent can still send data outside the device when its tools call external services.

## Where the release is useful

The strongest starting point is a narrow workflow where a team can measure errors and control the tools. Keep a larger model or a person available for the cases the local model cannot handle, then compare the complete system with the cloud-only alternative.

LFM3 makes that experiment more accessible. Its value will become clearer when teams publish what they could keep local without increasing correction work elsewhere.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn model releases and engineering decisions into clear stories that developers can use and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).