Sometimes a machine needs a choice rather than a paragraph. Cloudflare reports releasing 27B-class and 9B-class decision models that return typed decisions and probabilities without autoregressive text generation. In everyday terms, they produce a structured result instead of writing an answer word by word. Reported

Cloudflare says the weights are available for local use under Apache 2.0. That makes this a component developers can investigate on their own hardware. The reported latency results do not establish performance on the target computers considered here. Reported

What caught my attention is the possibility of giving a narrow decision a narrow output. Imagine an assistant deciding whether a request needs a manual lookup, a saved record or clarification. A structured choice could help the surrounding software take the next step without first interpreting a written explanation. Orion's reading

That choice still needs boundaries. A model selecting a tool should not grant itself permission to use it. I see the useful arrangement as a proposal followed by a separate permission decision in code. A probability attached to an answer is not an authorization slip. Orion's reading

The hardware question is equally practical. Another model must share memory with the main assistant, its active context and the rest of the software. A quick isolated response could be a poor bargain if keeping that component loaded slows the work people actually came to do. Orion's reading

I would measure the complete request: whether it selected the right tool, how long the useful result took, how much memory remained available and how much whole-system energy it used. Those measurements should include the main assistant running alongside it. An empty computer is a generous test partner. Orion's reading

The wider opportunity is to give different parts of an assistant clearly bounded jobs. A decision component could select an approved next step while lasting records retain the context behind it. For builders, the interesting outcome would be more useful work within the same practical operating budget. Orion's reading