- Published on
AWS Strands Labs Releases Strands Decider 2B: An Open Source Decision Model That Picks Options in About 115 ms
- Authors
- Name
- DeepHorus
Amazon Web Services' Strands Labs has released Strands Decider 2B, an open source decision model that refuses to write prose entirely. Instead of generating text, it reads a state plus typed questions and returns a choice, a yes/no probability, or a score, each paired with a calibrated confidence. The 1.9-billion-parameter model runs locally on a CPU, a consumer GPU, or an Apple silicon Mac, delivering median decisions in roughly 115 milliseconds. The release signals the arrival of a lightweight, deterministic alternative to general-purpose language models.
Decision models, sometimes called System One models, only became a named category after TypeSafe AI launched Jev last month. The distinction matters: a large language model can produce arbitrary output, while a decision model only selects among supplied options or rates on a fixed scale. Strands Labs positions its model as a fast, predictable complement to reasoning models rather than a replacement, explicitly warning that it is unsuited for coding, chat, or summarization tasks.
The model supports three question types: choice, which picks one of N options; noul, which returns a yes/no probability between 0 and 1; and score, which selects a level on an ordered rubric. Every answer is drawn from the allowed options. Weights ship on Hugging Face under Apache-2.0, and pip install strands-decider provides a CLI and an HTTP server, though that server binds to 127.0.0.1 without authentication, so production deployments must supply their own auth layer. No hosted inference provider serves it yet.
Architecturally, the team started from Qwen3.5-2B-Base and discarded the language-modelling head, replacing it with a pointer head of roughly one million parameters. That head compares the hidden state at the
position against each option's last-token hidden state, producing a result in a single forward pass with no decoding loop. The torso uses a rank-16 LoRA while the head runs in fp32, and the released checkpoint is v19. Because the state is read once, extra questions cost only their own tokens.
Benchmarks are measured on JevBench, a third-party evaluation suite for Jev-class models. Version 19 posts 0.723 public accuracy (167 of 231 tasks), a Brier score of 0.342, and an expected calibration error of 0.052. Tier accuracy runs 1.000 easy, 0.875 standard, and 0.505 hard. Latency is 115 ms median and 299 ms p95 on an RTX 3090, and 153 ms warm median on an M3 Pro for inputs under 300 tokens.
On the September 25 v1.4.2 board, v19 ranked third of 33 in the 2B class, and first of 30 after excluding three models sitting just above 2B. The repository honestly flags a caveat: Mapika's newer decider-2b v11 scores 175 of 231 on the Strands harness, eight tasks ahead of Strands Decider. The model was also absent from the newer v1.5.4 composite board at the time of writing.
Calibration is the practical payoff for builders. On unseen short classification tasks, answers marked at 0.9 confidence or higher were correct about 95% of the time, so the team advises confirming or escalating anything below that threshold. Early uses include model routing, tool selection, argument checking, triage, guardrails, evals, and hybrid agents where an LLM handles the hard calls while the decider manages rote ones.
The bundled example gates a weather tool call inside a Strands agent: two yes/no questions check whether arguments are grounded in the user's request and whether calling now is premature. From the CLI, routing a failing-payouts complaint across billing, sales, and retail returns billing with confidence 0.768. Expect community forks, hosted wrappers, and continued benchmark churn in the months ahead.
