Liquid AI Releases LFM3: Six Open-Weight 2B-Class Models for On-Device AI

Liquid AI has released LFM3, the third generation of its Liquid Foundation Models, comprising six open-weight 2B-class models built to run text, vision, audio, and decision tasks directly on-device. These models are now available on Hugging Face, offering developers new options for efficient, local AI applications. For broader context, explore our Top 100 AI Tools.

Introducing the LFM3 Model Family

The LFM3 family is built around a shared backbone, LFM3-1.7B-Base, from which all six specialized 2B-class models are derived. These models are engineered for on-device deployment, making them suitable for applications where local processing is critical. Each model in the LFM3 family can be fine-tuned using a single GPU, which could streamline development and deployment for various use cases. For broader context, explore our AI Tools Pricing.

Key Capabilities Across Modalities

The LFM3 series includes models tailored for specific AI tasks:

• LFM3-1.7B (Text): This text model operates within approximately 1 GB of memory at 4-bit precision and supports a 256K-token context window. It achieved a score of 38.42 on WebArena-Verified, a notable increase from its predecessor's 3.02. Liquid AI states that LFM3-1.7B outperforms Qwen3.5-9B on long tasks within browser and business applications.
• LFM3-VL-2B (Vision-Language): This model demonstrated an average score increase from 57.6 to 71.4 across 62 vision benchmarks. Liquid AI indicates this performance is comparable to Gemma 4 12B. A robotics demonstration using LFM3-VL-2B achieved a 96.2% average success rate across four LIBERO suites, generating action chunks in under 50 ms on a Jetson AGX Orin.
• LFM3-Audio-2B (Audio): This model offers speech synthesis at approximately 87 times real-time speed on a single CPU thread using an Apple M4 Max. It boasts a time to first token as low as 50 ms and supports 18 languages.
• LFM3-Omni-2B (Multimodal): This model reportedly surpasses other models with up to 10.7 billion parameters, including Qwen2.5-Omni-7B, on benchmarks such as Daily-Omni, WorldSense, and AV-SpeakerBench.

On-Device Performance and Accessibility

A core aspect of the LFM3 release is its focus on on-device performance. The models are designed to run efficiently on various hardware configurations, including devices like the Raspberry Pi 5 and those equipped with Qualcomm Hexagon NPUs. This emphasis on local execution can reduce latency, enhance privacy, and enable AI capabilities in environments with limited or no internet connectivity.

The open-weight nature of the LFM3 models, coupled with their availability on Hugging Face, makes them accessible for developers and researchers. This approach supports community-driven innovation and allows for custom adaptations and fine-tuning for specific applications.

Implications for Edge AI Development

The release of LFM3 could impact the development of edge AI applications. By providing powerful, yet compact, models capable of handling diverse tasks on-device, Liquid AI aims to lower the barriers for integrating advanced AI into embedded systems, robotics, and mobile applications. The ability to fine-tune these models on a single GPU further democratizes access to advanced model customization.

Conclusion

Liquid AI's LFM3 family represents a significant step in making advanced multimodal AI models more accessible and efficient for on-device deployment. With six open-weight, 2B-class models covering text, vision, audio, and decision tasks, LFM3 offers a versatile toolkit for developers. The reported performance benchmarks and the focus on local execution highlight the potential for these models to drive innovation in edge AI and specialized applications.

Sources

• Build specialized multimodal models with LFM3
• LiquidAI/LFM3-1.7B-Base · Hugging Face
• LiquidAI/LFM3-1.7B · Hugging Face
• LiquidAI/LFM3-VL-2B · Hugging Face