Description
LiquidAI LFM3-1.7B is a sophisticated text model developed for local agents and specialized applications. It is built on a tri-hybrid architecture, which includes 24 layers with a combination of gated short convolutions, sliding-window attention, and global attention. This model is post-trained for agentic harnesses like Hermes Agent, Pi, and OpenClaw, offering a 256K-token context window.
Read More
The model is designed to operate efficiently on edge devices, utilizing 4-bit precision to fit within approximately 1 GB of memory, making it compatible with most phones and laptops. It supports a wide range of languages, including English, Arabic, Chinese, French, German, and more, allowing for versatile applications across different linguistic contexts.
LFM3-1.7B is particularly suited for tasks requiring structured outputs and domain-specific fine-tuning. It can be adapted to various domains, formats, and tools through quick single-GPU fine-tuning. The model supports multiple inference frameworks and hardware configurations, including BF16, FP8, NVFP4, GGUF, MLX, and ONNX, providing flexibility in deployment.
Despite its capabilities, LFM3-1.7B is not recommended for agentic coding or complex multi-step reasoning due to its relatively small size. However, it excels in powering agentic applications that previously required larger models. The model's performance is validated through various benchmarks, demonstrating its efficiency in agentic applications, tool use, and instruction following.
For those interested in customizing the model, LFM3-1.7B offers fine-tuning support through methods like Supervised Fine-Tuning (SFT), Mixed Reinforcement Learning, and Multi-domain On-policy Distillation (MOPD). These training methods enhance the model's capabilities in conversation, reasoning, and tool-use tasks.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentCloud, Self-hosted
- PlatformsWeb
LiquidAI LFM3-1.7B Highlights
Tri-hybrid architecture
256K-token context window
4-bit precision for edge devices
Supports multiple languages
Post-trained for agentic harnesses
Domain-specific fine-tuning
Multiple inference frameworks
Efficient on phones and laptops
Getting Started with LiquidAI LFM3-1.7B
Access model page: Visit the LiquidAI LFM3-1.7B page on Hugging Face.
Authenticate: Log in to your Hugging Face account.
Configure: Select the appropriate inference framework and hardware.
Send prompts: Use the model for text generation tasks.
Fine-tune: Adapt the model for specific domains or applications.
LiquidAI LFM3-1.7B's Use Cases
- Local agents
- Specialized applications
- Tool calling
- Multilingual tasks
- Fine-tuning
