Now loading...
This year, Strands Labs unveiled a new venture aimed at those eager to explore cutting-edge techniques in agentic AI. Today, they are excited to introduce Strands Decider 2B, a compact decision-making model designed for rapid experimentation and local development.
The Strands Decider engages with a fresh category of decision models, also referred to as system one models, which have garnered increasing interest following the recent launch of Jev by TypeSafe AI. Unlike large language models that can generate diverse outputs, these decision models focus on selecting from predefined choices, such as determining whether a statement pertains to a specific object or identifying the language of a phrase, while providing simpler numerical assessments of options.
This focused approach enables decision models to run efficiently with low latency and to consistently yield definitive answers, albeit at the expense of flexibility and complexity in problem-solving compared to reasoning models. Additionally, decision models offer high-quality reliability scores for their predictions, a feature not found in traditional LLM inference APIs. This efficiency for multiple inquiries per prompt makes them particularly suitable for the agentic workflows being constructed with the Strands Harness SDK.
Strands Decider 2B is characterized by its two billion parameters, rendering it capable of functioning optimally on standard CPUs or GPUs while delivering responses in mere milliseconds. Its accuracy rivals that of existing models in this segment, and it has been made available as open source on GitHub, with model weights retrievable from Hugging Face, along with comprehensive training resources.
The model’s framework involves using a pre-trained large language model torso, specifically Qwen3.5-2B, with its language generation capabilities removed. This modification includes a small pointer head, which evaluates potential answers by comparing states in the model. This latest release follows iterative improvements, with project documentation outlining changes and offering insights into the development process.
Strands Decider 2B has been tested against key performance indicators: accuracy, calibration of confidence scores, and decision latency. Its performance places it competitively within its class, achieving third place in overall accuracy on JevBench’s public dataset. The model is also designed for swift local decision-making, achieving a median latency of approximately 115 milliseconds on typical hardware, with predictions becoming incrementally longer relative to task size.
The rationale for choosing a two billion parameter size aligns with Strands Labs’ goals of fostering experimentation. The model is intended to be run on accessible hardware, enabling a wide range of applications without significant overhead. It is well-suited for simpler tasks documented on JevBench, providing a robust platform for various projects.
Early adopters of Strands Decider 2B have found its utility in diverse applications including model routing, tool selection, evaluation processes, and various classification tasks. Its effective integration with large language models offers a promising path for building hybrid agents that can optimize decision-making processes while streamlining costs and enhancing efficiency.
To begin exploring Strands Decider 2B, users can easily start with the strands-decider command line interface. This CLI allows for simpler interaction with the model, making it easy to pose decision-based queries and receive swift responses with confidence metrics for guidance.
The Strands team is actively expanding resources for integrating decision models, encouraging the community to collaborate and innovate. Those interested can download and use the strands-decider-2b model today, exploring the wealth of accompanying data and starting points for development. Dive into the project on GitHub and access model snapshots through Hugging Face. Now is the time to start experimenting!
