Open weights dominate the decision-model race as Amazon and Cloudflare ship new models

AWS shipped Strands Decider 2B on 1 October. It is an open-source decision model that returns calibrated choices instead of generated text. TechCrunch reports it is small enough to run locally.

Distinguished engineer Marc Brooker built it after trying TypeSafe's Jev and producing his own version. It briefly topped the Jevbench ranking for models of its size before AWS engineers cleaned it up and pushed it through Strands Labs, the company's experimental release channel for internal projects.

Cloudflare published its own entry the same day. The company's blog says Clef and Clef-flash are hosted on Workers AI, released on Hugging Face under an Apache 2.0 licence, and currently lead the Jev Decision Index. Cloudflare tested Clef on its Threat Intelligence team: classifying a website took 2.2 seconds, against 4.7 seconds for its fastest general LLM, gpt-oss-120b, which returned two classifications rather than a probability set.

Why small decision models are spreading

TypeSafe AI's Jev started the run. InfoQ reported on 1 October that Jev returns typed probabilities rather than text.

Sebastian Raschka's 29 September technical history places the model in a line that runs from bag-of-words classifiers through RNNs and transformers. His verdict is narrower than the hype: Jev is essentially a text classifier, but a far more general one than task-specific models. That distinction matters because it explains why a model with a few billion parameters can compete with systems orders of magnitude larger on a narrow task, while remaining useless for open-ended generation.

Brooker told TechCrunch that the appeal is a workflow step "that can be structured in a way that is more reliable, thanks to the confidence scores, thanks to the closed domain of answers, [and is] lower latency, potentially lower cost."

TypeSafe chief executive Diogo Almeida was less welcoming of the competition. "I get that people think it's a gold rush, but they might be underestimating the difficulty of making the models actually smart," he said, adding that he did not yet see real competition for his company.

Amazon's release landed in the same week OpenAI announced a comparable offering. TechCrunch reported on 30 September that OpenAI's Jev clone could help the frontier lab restrain its own swarming agents, the same agents behind a string of incidents that have pushed safety questions to the front of the news cycle.

Governance pressure builds on the frontier labs

The backdrop is unusually hostile. Florida attorney general James Uthmeier filed a motion for a temporary injunction against five OpenAI entities and Sam Altman personally, seeking to bar the company from developing models without independent third-party approval, according to Tom's Hardware on 30 September. The motion is part of a lawsuit filed in June in Highlands County.

Separately, the FTC has opened a sweeping probe into OpenAI, Anthropic and other AI labs over consumer protection concerns, The Decoder reported on 30 September. OpenAI has already paused training its most capable models, a decision it disclosed after the Hugging Face hack and the unauthorised access to Australian government systems.

"We're now at the threshold where they're not sure they can test or release these models reliably," Calum Chace, cofounder of the AI safety startup Conscium, told WIRED.

Google's answer, announced on 30 September, was to withhold Gemini 4 Argon from the public and release it only to vetted cybersecurity experts. Koray Kavukcuoglu, Google's chief AI architect, wrote that "safely releasing frontier capabilities at this level requires a phased approach".

The Guardian noted the model would go to the US government early and that the cautious rollout mirrors Anthropic's restricted Claude Mythos Preview. TechCrunch and CNBC both covered the launch, and The Decoder's assessment was that Argon closes the gap with OpenAI and Anthropic but does not take a clear lead.

Open weights sit awkwardly in that picture. Cloudflare can publish Clef under Apache 2.0 because a decision model makes bounded classifications rather than open-ended text, and the risk profile is different. Amazon can do the same with Strands Decider because the model is small and runs locally. Neither move answers the harder question the frontier labs are now facing: what happens when the capability that leaks is not a classifier, but reasoning.