Decision models took 1 in 4 Vercel router requests in October
A decision model is a fast LLM that writes no text and selects from alternatives that are set before, and it is new.
In October, more than 1 in 4 requests on the Vercel AI model router went to decision models, not LLMs. An LLM is a large language model that writes text. A decision model writes none: it selects from alternatives that are set before and tells how sure it is. On Tuesday Microsoft told that Microsoft-Decision-1.1 is available for wide use on its cloud. The UK startup Sqwish raised $3 million.
How it works
A decision model makes a fast, cheap selection. This helps when speed is more important than the reasoning of a slow frontier model. The TypeSafe Jev family has the most use. OpenRouter handled more than 1.5 billion queries to TypeSafe models last week.
The proof
It is not easy to rank these models. Sqwish tells that its kettei fat model gave good results on some benchmarks. Builders tell that the correct test is to put in a model and look at the result. A Microsoft vice president tells that models tuned only for benchmarks do not work in use by customers.
What is not known
Sqwish tells that customers can continue to make their model better with their own production data. It gave no test results for this. No clear ranking is known.
Insight: Decision models are easy to copy, so your data is the edge
One Amazon engineer made a decision model in a weekend, and Microsoft, vLLM and others followed in weeks. A test of typed decision models found a hosted model that is sensitive to wording and underconfident. The model looks like a commodity. Because of this, the company that tunes it on your production data can keep you. Sqwish sells this. Keep your data and your test set.
Sources
Posted