DEVOURED
Liquid AI Launches LFM3 Models for On Device AI
Liquid AI's new LFM3 model family achieves strong agentic performance on sub-2B parameter footprints, outperforming much larger models on edge hardware.
What: Liquid AI released the LFM3 family of 2B-parameter models featuring text, vision, audio, and decision-making capabilities. The 1.7B text model supports a 256K context window and fits in roughly 1 GB of memory at 4-bit precision, while the vision-language variant matches Gemma 4 12B in benchmarks.
Why it matters: This demonstrates a shift toward highly optimized, domain-specific 'small' models that can perform complex agentic tasks locally, bypassing the latency and privacy concerns of cloud-based APIs.
Takeaway: Developers can fine-tune these models locally on a single GPU using the open weights now available on Hugging Face.
Deep dive
- Shared backbone: All LFM3 variants share a tri-hybrid architecture of gated convolutions, sliding-window attention, and global attention.
- Hardware-in-the-loop: The architecture was optimized using 14 device combinations including Apple M4 Max, Raspberry Pi 5, and Qualcomm NPUs.
- Efficiency: 1.7B text model uses 35% fewer parameters than its 2.6B predecessor but delivers superior performance.
- Agentic Post-training: Models underwent multi-domain on-policy distillation (MOPD) and reinforcement learning to handle tool-use and multi-step reasoning.
- Video Support: The LFM3-VL-2B model uses a new vision encoder that is 3x more compute-efficient than the previous SigLIP2-based approach.
- Audio/Speech: Includes a proprietary tokenizer, LiquidCodec, for multilingual speech generation at faster-than-real-time speeds on a single CPU thread.
Decoder
- Quantization: Reducing the precision of model weights (e.g., from 16-bit to 4-bit) to decrease memory usage and improve inference speed.
- Tri-hybrid architecture: A neural network structure combining convolution layers for local pattern recognition with attention mechanisms for long-range sequence modeling.
- KV cache: A memory buffer that stores previously computed key and value states during inference to avoid redundant calculations in transformer models.
- SGLang: A structured generation engine optimized for high-throughput LLM inference.
Original article
Full article content is not available for inline reading.
Explore all 54 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi