Key Findings
The LFM3 family consists of six models derived from the LFM3-1.7B-Base. The core text model (1.73B parameters) supports 16 languages with a 256K token context window. Notably, the LFM3-VL-2B vision model achieves performance comparable to the larger Gemma 4 12B, while the LFM3-Omni-2B outperforms models up to 10.7B parameters, including Qwen2.5-Omni-7B, in audio and visual understanding benchmarks. The d1-2B decision model improves ImajevBench scores to 75.5 from 64.0 compared to d1-3B.
Technical Details
The LFM3-1.7B-Base architecture features 24 layers: 13 gated short-convolution layers (kernel size 7), 5 sliding window attention layers (256 tokens), and 6 global attention layers. Technical optimizations include:
- Gated Attention: Sigmoid gates suppress activation outliers to improve quantization.
- Partial RoPE: Applied fully to sliding window layers and only to the high-frequency half in global attention layers.
- Sandwich Normalization: Normalizes both input and output of each block.
The model was pre-trained on 22T tokens, with the final 3.5T tokens focusing on agentic execution and tool calling. Post-training involved four stages: SFT (1.75T tokens), mixed RL, on-policy distillation via six teacher models, and agent-based RL.
Hardware and Integration
Liquid AI conducted architecture searches across 14 hardware/backend combinations, including Samsung Galaxy S26 Ultra, Raspberry Pi 5, Apple M4 Max, and NVIDIA B300. On a single B300 GPU, LFM3-1.7B can generate 3.6 billion tokens per day with 1,024 concurrent requests. The models natively support llama.cpp, MLX, and vLLM, with weights available in BF16, FP8, NVFP4, GGUF, MLX, and ONNX formats.
Significance and Outlook
LFM3 demonstrates significant efficiency gains; the 4-bit text model requires approximately 1 GB of memory. In robotics, the vision-language-action model achieved a 96.2% average success rate across four LIBERO suites and generates action chunks in under 50ms on Jetson AGX Orin. Launch partners include AMD, Arm, NVIDIA, Qualcomm, and Intel.
Source: https://www.unite.ai/ja/liquid-ai-launches-lfm3-family-of-six-open-weight-2b-class-models/
Get our weekly technology intelligence — free
Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.
Subscribe Free — Weekly Tech Intelligence
By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.
- Your email and selected fields are used only to deliver the newsletter.
- We never share your information with third parties.
- You can unsubscribe anytime via the link in each email.
See our Privacy Policy for details.
Takes about a minute · Unsubscribe anytime
