Strands Decider Gemma 4

// Overview

Strands Decider Gemma 4 is a family of four open decision models that AWS's Strands Labs published on Hugging Face on October 9, 2026 and announced on the Strands Agents blog on October 15, 2026. The four sizes, strands-decider-E2B-gemma4-v1-2610, E4B, 12B and 26B-A4B, are the first Strands deciders built on Google's Gemma 4 (on gemma-4-E2B-it, E4B-it, 12B-it and 26B-A4B-it), following Strands Decider 2B on Qwen3.5-2B-Base from October 1, 2026. Like that model, they do not generate text: they pick between options or rate something on a scale and return a confidence with every answer, for the decisions inside agentic workflows such as routing, checking tool arguments, guardrails and evaluations. All four accept the same three question types, noul (yes/no), choice (one of N) and score (a level on an ordered scale), so a workflow can switch sizes without rewriting its questions.

Each size ships its own LoRA adapter on the Gemma 4 text decoder and a small readout head; the loader fetches the matching Gemma 4 base weights separately. Every checkpoint is a soup of three training runs with seeds 0, 1 and 2: each layer's LoRA update is the exact mean of the three seeds' updates, stored as one rank-48 adapter, and an option's score is the mean of the three heads' scores, after which the soup is calibrated with one temperature per question type. Each seed trained for one epoch of 4,109 steps on a host with eight NVIDIA H100 80GB GPUs, on 123,339 rows of the Hobson v21 data recipe plus 12,263 code-task rows, using Qwen3.5-4B's output distributions as targets where that teacher agreed with the gold label. The same day, Strands Labs also published strands-decider-2B-qwen3.5-v1-2610, a retrained Qwen3.5-2B decider under the new naming scheme strands-decider-<size>-<base>-v<N>-<YYMM>.

On JevBench v1's 231 public tasks with a 4,096-token window, the four sizes answer 180 (77.9%), 187 (81.0%), 200 (86.6%) and 207 (89.6%) tasks correctly, with Brier scores falling from 0.313 to 0.162. The launch post reports median latencies per question on one NVIDIA H100 80 GB of 102 ms (E2B), 139 ms (E4B), 136 ms (12B) and 245 ms (26B-A4B), and notes that the 12B scored about three points below the 26B-A4B with roughly 45% lower median latency in that run. Strands Labs recommends starting with E2B and moving up only if it misses the target. The README lists approximate bf16 GPU memory of 9 GB, 15 GB, 23 GB and 51 GB for the four sizes.

// Benchmarks

This model's scores

1. JevBench v1, public set (231 tasks), 26B-A4B89.6%
2. JevBench v1, public set (231 tasks), 12B86.6%
3. JevBench v1, public set (231 tasks), E4B81%
4. JevBench v1, public set (231 tasks), E2B77.9%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• 207 of JevBench v1's 231 public tasks (89.6%) for the 26B-A4B, with a Brier score of 0.162 and an expected calibration error of 0.051
• 200 of 231 (86.6%) for the 12B at a 136 ms median per question on one H100, against 245 ms for the 26B-A4B
• The same noul, choice and score questions work across all four sizes, so a model can be swapped by changing its ID
• 48.69 on Decision Index 0.2.1 (balanced skill) for the 26B-A4B and 47.14 for the 12B
• Apache-2.0 adapters, readout heads and code, with the training recipe in the public repository

// Best for

• Reach for it to route requests, check tool arguments or apply guardrails inside a Strands agent when the 2B decider is not accurate enough for the task.
• Reach for the 12B when accuracy matters but latency is tight: in the launch run it scored about three points below the 26B-A4B at roughly 45% lower median latency.
• Reach for it to score or classify outputs in an evaluation pipeline, acting only on answers above a confidence threshold tested on your own traffic.
• Look elsewhere for long, multi-step documents, which the model cards list as a weak spot, or for anything that needs generated text.

// How to access

// Strands Decider (open decision models) — every version

The full lineage of the Strands Decider (open decision models) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.