AWS releases a model that can't write, and that's the point

On 1 October, AWS's Strands Agents project released Strands Decider 2B, a model of roughly two billion parameters, and the most interesting thing about it is what was taken away. According to the Hugging Face model card and the GitHub repository, it starts from Alibaba's Qwen3.5-2B-Base with the next-token prediction head removed and replaced by a pointer head that scores the options it is given: just over a million new parameters, adapted with a rank-16 LoRA. "Decision models are designed to pick between sets of options... and assign simple numerical scores", says the official blog post by Marc Brooker, Mike Chambers and Fabio Nonato de Paula. This isn't a crippled chatbot. It's a component built for a class of tasks that shows up everywhere in agents: which tool to call, which model to route a request to, whether an action stays within bounds. Choices between alternatives that are already written, where making an LLM generate text is wasted work.

The claimed numbers: 72.3% on the public JevBench set, i.e. 167 of 231 tasks; 87.2% on ContractNLI, 88.4% on MuSiQue, 71.7% on HotpotQA. Median latency of about 115 ms on an RTX 3090 and 153 ms on an M3 MacBook, growing roughly linearly with task size. The blog ranks the model third of 33 in the 2B class, first of 30 if you exclude models just above the size threshold. It has to be said that the sources don't line up here: VentureBeat reports a 106 ms median and second place, but it measured v18 with the HTTP round trip included and 7.7 seconds of server warm-up discarded, whereas the blog and the repository refer to v19. The latency gap may come down to version and measurement method. The ranking gap remains unexplained: that's an open question worth keeping open.

The list of limitations is signed by AWS itself, and it's worth reading: no code, no chatbot, no summaries; less capable than reasoning models on complex problems; weak on long, multi-hop documents; calibration tuned only on short classification tasks; label noise inherited from public datasets. Above all, the model's judgement should not be treated as a security boundary against adversarial input — which is exactly the temptation, given that guardrails are among the suggested uses. The natural comparison is with TypeSafe's Jev, a proprietary decision model offered as a hosted service, which we've already covered. According to VentureBeat, the AWS materials contain no head-to-head accuracy comparison with Jev. Nor is there any evidence that running it yourself costs less than the hosted service. The dedicated integration libraries are, for now, still in development.

What sets this release apart isn't the weights themselves but the weights together with the recipe: code, training procedure, data inventory, evaluations, Apache 2.0 licence. The performance is still self-reported and no independent check exists yet, but with the recipe out in the open anyone can try to confirm it or disprove it — which is a different situation from trusting a number on a product page. What strikes me is that the most solid contribution here is a subtraction: taking away a model's ability to talk to see how well it does the one thing it was asked to do.

— Olya