Strands Decider 2B: an open model that returns decisions, not text

Amazon has released Strands Decider 2B, an open model that scores predefined answer options instead of generating text. The weights and training code landed on Hugging Face on 1 October 2026 under Apache 2.0, with no hosted API. It is the first free counterpart to TypeSafe AI's closed Jev.

Key takeaways

• The base is Qwen3.5-2B-Base, with a rank-16 LoRA adapter adding just over a million parameters
• JevBench, public split: 167 correct answers out of 231 tasks, or 72.3%
• Brier score 0.348 and expected calibration error 0.050 — the model returns probabilities, not bare labels
• Median latency of 106 ms on an RTX 3090, 296 ms at the 95th percentile
• Training took roughly 70 minutes on eight H100s, or 11 hours on a single RTX 3090

A pointer head instead of next-token prediction

Decider strips the next-token layer off Qwen3.5 and replaces it with a small head that scores every allowed answer option. The agent gets a probability distribution over the candidate list rather than a sentence to parse.

With it goes the whole intermediate layer that wraps every language-model call today: forcing JSON, validating the schema, re-asking for a valid format.

TypeSafe AI demonstrated the same idea in September with Jev, but that model stayed closed and API-only. Amazon shipped the weights, the data and the training scripts, so the result can be reproduced on local hardware.

The numbers, without the marketing

An accuracy of 72.3% is useful but far from certain — one decision in four is wrong. Calibration is the more interesting figure.

What a Brier score is, and why it matters here

The Brier score^?Brier score: The mean squared gap between a stated probability and what actually happened. Zero is perfect, 0.25 is coin-flip territory. measures not accuracy so much as the honesty of a model's confidence.

A calibration error of 0.050 means the model's stated probability tracks reality, so a cutoff threshold can be set deliberately. On narrow tasks the results are clearly better — 0.884 on MuSiQue and 0.872 on ContractNLI.

{
  "jevbench_public":   { "accuracy": 0.723, "correct": "167/231" },
  "calibration":       { "brier": 0.348, "ece": 0.050 },
  "musique":           { "accuracy": 0.884, "n": 1199 },
  "contractnli":       { "accuracy": 0.872, "n": 1026 },
  "held_out_short":    { "accuracy": 0.641, "n": 6000 }
}

Cost matters as much as quality here. A 106 ms median on a single RTX 3090 means decisions happen locally, inside the agent loop, with no external API call.

Where Decider stands against its rivals

Why it matters

Agents spend a lot of time on small rulings: whether to call a tool, whether a document is relevant, whether to close the loop. Handing those to a large text-generating model is expensive and unreliable. An open, calibrated 2B model changes the economics of that layer and undercuts the case for a paid API built around it.

What's next

• Amazon launched no hosted API, so deployment requires in-house infrastructure
• The 72.3% figure covers the public JevBench split, the private split has not been released
• The hobson-v19 variant name on the model card points to further adapter iterations

Sources

• VentureBeat — Amazon unveils a free, fast, open source Jev killer: Strands Decider 2B makes decisions in fractions of a second
• Hugging Face — StrandsAgents/strands-decider-2B-hobson-v19
• Strands Agents — Strands Agents SDK