AWS team releases Strands Decider 2B, an open decision model that runs on CPU

Strands Labs, the GitHub organization where AWS gathers experimental projects around its open-source agent SDK Strands Agents, released Strands Decider 2B on October 1, 2026. According to the project page, it is the first member of a new family of decision models. It is a small model with about 1.9 billion parameters, and both the code and the open weights are available under the Apache 2.0 license. At this size, it can run locally on your own computer or server without a powerful GPU.

A small model that picks between options

A decision model picks from a predefined set of options for a given text, or rates it on a scale. The project page places these models between large language models and traditional classifiers: they answer faster than an LLM, and they do not need to be trained separately for each task the way a traditional classifier does. Strands Decider answers three types of questions: yes/no questions, picking one of several options, and scoring on an ordered scale. For a customer message such as "My payouts have been failing for 3 days," a single command can ask which team should handle it, whether it is urgent and how frustrated the writer is.

Every decision comes with a confidence score. According to the project page, on short classification tasks it has not seen before, the model's high-confidence answers are right in the large majority of cases; when confidence is lower, it recommends confirming the answer or asking a person. The suggested uses are routine decisions inside agent workflows: picking the right LLM for a task, choosing an agent's next tool, checking a tool call's arguments before it runs, routing incoming requests to the right team, safety and policy checks, and scoring model outputs at low cost. The project proposes hybrid agents that leave hard decisions to the LLM and routine ones to the decider.

Built on Qwen, with the text-generation head removed

The model is built on Alibaba's Qwen3.5-2B-Base. The developers removed its language-modeling head and replaced it with a small pointer head of about a million parameters. That head scores each option by comparing it against the model's internal representation after reading the text, so the answer comes out in a single forward pass with no text generation. Because the options are not baked into the weights, they can be redefined with every request and there is no cap on how many a question can have. When several questions are asked about the same text, the text is read once and each extra question adds only its own tokens.

The team measures accuracy and how well the confidence scores hold up on JevBench, a third-party benchmark for this class of model. The research record is public in the repository: predictions are written down before each training run, the outcome is appended afterwards, and runs that miss their bar stay in the record.

Installation and hardware

The model installs with pip install strands-decider. The bundled command-line tool lets you ask questions directly, and strands-decider serve runs the model as an HTTP server that other applications can call. The weights are published on Hugging Face under the StrandsAgents account. According to the project page, the model runs on NVIDIA GPUs, on Apple silicon Macs and on CPU-only machines, and faster MLX support for Macs is on the way. Retraining from scratch is also possible: the full recipe takes about 11 hours on a single RTX 3090. The repository includes an example inside a Strands agent, where the agent checks two yes/no decisions before calling a weather tool and asks the user for the city instead of guessing when it is unclear.

Decision models are emerging as a category of their own in recent weeks. After Typesafe AI's Jev, Cloudflare released its Clef decision models as open models on the same day. Strands Decider's server also answers on the same /v1/systemone endpoint name used by Jev and Clef.

All the published measurements were run by the project team, which itself notes that on a benchmark with relatively few tasks, small differences can fall within the variation between retrains. The pages do not say which languages the model supports, and its training data sources are mostly English datasets, so it is worth testing it on your own examples before using it on other languages. The amount of text it can read at once is limited to a few thousand tokens, and longer texts are truncated. For a fuller picture, it is worth following evaluations from independent researchers.

Sources: GitHub: Strands Decider, Hugging Face: Strands Decider 2B, AWS: Introducing Strands Labs.