Strands Decider 2B
Released Oct 1, 2026 · knowledge cutoff unpublished
- Status
- Unknown
- Location
- United States
- Type
- System One
- Modality
- Text
- Context window
- —
- Max output
- —
- Speed
- —
- Price ($/MTok in / out)
- — / —
- Cost per task
- —
- Open weights
- Yes
Strengths & weaknesses
Strengths
- Small enough to run on a laptop or a single consumer GPU, answering a typed question in a little over a tenth of a second at the median on an RTX 3090
- The whole recipe is open — training data inventory, scripts and evaluations ship alongside the weights, so the model can be retrained rather than just fine-tuned
- Swaps the language-model head for a pointer head of about a million parameters that scores each option directly, so there is no generated text to parse
Weaknesses
- Calibration is a single temperature per question type fitted on short classification tasks — AWS's own card says to measure confidence on your own traffic before trusting a threshold
- Tail latency is uneven off a discrete GPU: AWS measures a 95th percentile in the seconds on an M3 Pro, against a sub-quarter-second median
- Behaviour is a moving target — the launch checkpoint was hobson-v19 and AWS had already published hobson-v21 four days later
Not an LLM: a "System One" decision model from AWS's Strands Labs, a LoRA adapter on Qwen3.5-2B-Base plus a roughly 1M-parameter pointer head that scores the options of typed noul, choice and score questions, so none of this site's benchmarks apply and every cell is null by design. Filed under the fast tier alongside Jev and Laya. Released on Hugging Face (StrandsAgents/strands-decider-2B-hobson-v19) and GitHub (strands-labs/strands-decider) on 2026-10-01 under Apache 2.0; v21 followed on 2026-10-05. Availability is self-host: AWS offers no hosted endpoint or price, and the community ONNX and WebGPU ports are not a hosted service. AWS evaluates it with a 4,096-token context but publishes no maximum, so contextWindow is null. AWS reports 0.723 (167/231) on JevBench public, with medians of 115 ms on an RTX 3090 (p95 299 ms) and 234 ms warm on an M3 Pro (p95 2,628 ms); all are its own figures, not independently reproduced. Training drew on ContractNLI, MuSiQue, synthetic rows generated by open-weight models, and distillation from Qwen3.5-4B. The v21 card documents an opt-in vision mode that keeps the base model's vision tower without any image training; modality stays text for the launch checkpoint. No Artificial Analysis coverage; knowledge cutoff unpublished.
Benchmarks0 of 10 reported
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
Terminal-Bench 2.1
AIME (latest)retired
Humanity's Last Exam
LMArena Elo
GDPval-AA v2
ARC-AGI-2
For developers
API model strings
Not researched
Licence
Apache License 2.0 · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor