Small language models are often treated as chatbots with fewer parameters; this one is aimed at a different bottleneck: running useful agents where memory and inference speed matter. Its hybrid design combines short convolutions with local and global attention, pairing a 1.73B-parameter footprint with a 256K-token context window.
Key Capabilities
- Agent and tool workflows: Post-training targets agent harnesses and tool use, making it better suited to local task execution than a general-purpose small chat model.
- Long-context edge inference: The 256K context supports extended agent histories, while 4-bit weights need about 1 GB of memory—useful for phones and laptops.
- Adaptable deployment: Quantized variants support several inference stacks, and TRL-based fine-tuning lets teams tailor behavior to their own data and tools.
- A hybrid backbone: Its 24 layers combine gated short convolutions, sliding-window attention, and global attention to balance efficient local processing with access to distant context.
Who It Suits—and the Trade-off
A good fit for developers building on-device assistants, tool-using agents, or narrow domain applications that benefit from customization. Look elsewhere when the main requirement is strong complex reasoning, agentic coding, or answering knowledge-heavy questions without retrieval: the model card explicitly recommends retrieval for knowledge-intensive tasks and cautions against relying on it for complex reasoning.
