What happened
Liquid AI released the LFM3 model family on October 27, comprising six models for language, vision, speech, and structured decisions. The models share a 1.7-billion-parameter foundation, with specialized versions around 2 billion parameters. The release includes a 256K-token context window for the text model, a tri-hybrid architecture combining convolution and attention, and specific pricing for hosted API access. The company reports that the smallest text model outperforms larger rivals on selected browser automation and tool-use tasks, while acknowledging limitations in agentic coding and complex reasoning.
Liquid AI introduced the LFM3 family on October 27, consisting of six models spanning language, vision, speech, and structured decisions. The core text model, LFM3-1.7B, supports a 256K-token context window, though the hosted API version is limited to 128,000 tokens. The models are built on a tri-hybrid architecture that combines convolution layers for efficient local processing with attention mechanisms for broader context awareness.
The release emphasizes enterprise control, allowing customization on a single GPU and deployment on private servers or cloud infrastructure. Liquid reports that the 1.7B text model scores 38.42 on WebArena-Verified, significantly outperforming Qwen3.5-9B (28.94) and its predecessor LFM2.5-2.6B (3.02). However, the company acknowledges that larger models still outperform LFM3 in agentic coding and knowledge-intensive tasks.
Pricing for the hosted service is available in the Liquid console, with the text model costing $0.08 per million input tokens and $0.20 per million output tokens. The vision and audio models have higher per-token costs. For local deployment, the models are available under the LFM Open License v1.0, which requires a separate commercial agreement for businesses with annual revenues exceeding $10 million.
The family includes LFM3-VL-2B for vision tasks, which adds video input and reports an average score of 71.4 across 62 evaluations. The LFM3-Omni-2B model combines vision, speech, and language, capable of interpreting video with soundtracks or combining spoken instructions with images. The d1-2B decision model offers structured outputs for classification and extraction tasks, providing probability scores for automated decision-making.
Source details: venturebeat.com
Why it matters
This release provides enterprises with a viable option to run AI agents on local, company-controlled hardware, reducing dependence on remote AI services and enhancing data privacy. The compact size, fitting into approximately 1GB at four-bit precision, allows deployment on devices like laptops and mobile phones, enabling offline or low-latency workflows. The specific pricing and licensing terms, including a revenue threshold for commercial use, offer a clear cost structure for organizations seeking to integrate AI into repeatable business processes without incurring high per-token costs for every interaction.
The primary value of LFM3 lies in its ability to run on enterprise-controlled hardware, which addresses privacy concerns and reduces latency for sensitive data. The compact size, requiring only about 1GB of memory at four-bit precision, makes it suitable for deployment on edge devices, such as laptops and mobile phones, where traditional large models are impractical.
The specific benchmark results, particularly in browser automation and tool use, suggest that LFM3 can handle defined, repeatable workflows effectively. This is relevant for companies looking to automate tasks like inbox organization, scheduling, and web research without relying on external AI services that may have data retention policies or higher costs.
The licensing terms introduce a clear barrier for larger enterprises, requiring a separate agreement for commercial use if annual revenue exceeds $10 million. This structure allows smaller organizations to adopt the technology more easily while ensuring Liquid AI captures value from larger deployments. The policy for the hosted service further enhances its appeal for privacy-sensitive buyers.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?
What to watch next
Enterprises should monitor the real-world performance of LFM3 in specific workflows, such as document extraction and ticket classification, to validate the company-reported benchmarks. The adoption of the LFM Open License v1.0, particularly for companies exceeding $10 million in annual revenue, will determine the model's accessibility for larger organizations. Additionally, the integration of LFM3 into existing agent frameworks like OpenClaw and Hermes Agent will be crucial for assessing its practical utility in complex, multi-step tasks.
Independent validation of the benchmark results is crucial, as the current data is company-reported. Enterprises should conduct their own testing on specific workflows to assess accuracy, latency, and reliability, particularly for tasks involving failed tool calls and retries.
The adoption of LFM3 by major partners like Mercedes-Benz and Shopify, who are listed as pre-release testing partners, will provide insights into its real-world performance. However, no confirmed production rollouts or measured business outcomes have been disclosed yet.
The evolution of the LFM Open License v1.0 and its impact on the broader AI ecosystem will be important to monitor. The revenue threshold for commercial use may influence how widely the model is adopted across different sectors, particularly in industries with strict data governance requirements.