AWS published a 2B decision model that runs on one consumer GPU

On 2026-10-01, AWS Strands Labs (part of Amazon Web Services) released Strands Decider 2B under the Apache 2.0 license. Strands Decider is a small "decision model" (a type of system-one/model that selects between options rather than generating free-form text) built by fine-tuning Qwen/Qwen3.5-2B-Base with LoRA and adding a small pointer head. The model is available as open weights, together with code, training scripts/configs, evaluation data, and documentation via the public repositories. The release provides both the model checkpoint and usage tooling (CLI and server).

AWS entered the public space of open-source "decision models" (structural/option-picking models for agent control loops) with a concrete 2B implementation alongside published training/eval assets and a permissive license. The model is positioned as returning decisions (choices/scores) rather than prose, with calibrated confidence scores per answer.

Before: Decision logic in agent loops was often handled by large LLMs generating text and then parsing, or by traditional classifiers requiring bespoke training; open, lightweight decision models were an emerging class with limited public availability from hyperscalers.

Change: AWS Strands Labs released Strands Decider 2B (Apache 2.0) with weights, training data/scripts/configs, evaluation artifacts, and documentation, reporting performance metrics on JevBench v19.

After: A 2B Apache-2.0 decision model fine-tuned from Qwen3.5-2B is publicly available for local/edge use, enabling developers to replace some agent routing/triage/gating decisions with a small, fast model that returns structured decisions with confidences.

Strands Decider uses a decoder LLM torso (Qwen3.5-2B-Base) with the language modeling head removed and replaced by a small pointer head (~1M parameters). A rank-16 LoRA adapter is applied to the torso; the head runs in fp32. For a given state plus question options, it scores each option (one forward pass, no generation/decoding loop) and returns the selected choice/score with a confidence. The model supports three question types from the same mechanism: noul (yes/no), choice (pick one of N), and score (ordered rubric scoring). It is natively multimodal-capable via the Qwen3.5-2B-Base vision support in the broader ecosystem (vision use described in docs), and designed so multiple questions about the same state can be answered efficiently in one pass context. (Architecture described in GitHub repo/docs; weights and code published.)

This release is part of a broader pattern where hyperscalers are open-sourcing small, specialized models for agent control loops. Combined with S34 (Cloudflare Clef/Clef-flash) in the same week, it signals that decision-model architectures are crossing into mainstream infrastructure. A 2B parameter model that runs well on consumer GPUs makes local/private workflow routing plausible for regulated/edge deployments.

• Local/private decision-making: Running a 2B decision model on a single consumer GPU (e.g. RTX 3090) makes on-device/edge routing/gating feasible without sending sensitive state to remote LLMs.
• Lower cost/latency for rote decisions: Separate "system one" decisions from "system two" reasoning; route simple triage/tool-selection to a small model while reserving larger LLMs for complex reasoning.
• Transparent/reproducible research: Apache 2.0 weights plus training scripts/configs/data inventory and evaluation artifacts enable reproducibility and community iteration.

• How does Strands Decider 2B perform on real enterprise agent routing tasks vs. JevBench?
• Is the training data mix optimal for production triage/routing in regulated domains?
• How does calibration hold up under distribution shift in production?
• Are there significant differences vs. Cloudflare Clef/Clef-flash (S34) in calibration/accuracy/latency characteristics?

1. Community adoption: Developers will integrate into Strands Agents SDK and other agent frameworks; community benchmarks may emerge.
2. Independent evaluation: Third parties may reproduce JevBench results and evaluate on additional domains.
3. Ecosystem convergence: Comparison with Cloudflare Clef/Clef-flash (S34) will clarify trade-offs (size, latency, calibration, context).
4. Iterative improvement: Model variants/fine-tunes may appear based on community feedback and additional training data.

AWS's release of Strands Decider 2B marks a concrete step toward specialized "decision models" becoming standard infrastructure in agentic AI. Built on Qwen3.5-2B with Apache 2.0 licensing and published training/eval assets, it enables local, fast, structured decision-making at 2B scale. Combined with a similar Cloudflare release the same week, this suggests the industry is converging on splitting agent loops between lightweight decision models and larger reasoning models. The reported performance (~72% on JevBench v19, ~106ms on RTX 3090) is a COMPANY CLAIM pending independent verification; the architectural approach and openness are the more durable signals.