Skip to content
SurenoGet Sureno Browser
← Sureno News

Decision model use skyrockets as customization emerges as the next frontier | TechCrunch

A fast-growing model category

In October, decision models received more than one in four requests on Vercel’s AI model router. The category had emerged only a month earlier, according to the report.

Unlike models that generate output tokens, decision models choose among preset options and register confidence in their choices. They can suit production settings where speed and cost matter more than the reasoning of a slower, more expensive frontier model.

Big companies join the race

TypeSafe became a $7.5 billion company after launching its Jev family, which continues to lead usage in visible measures. OpenRouter handled more than 1.5 billion queries to TypeSafe’s models last week.

Many alternatives followed Jev’s release. An Amazon engineer built one over a weekend, hyperscalers introduced their own offerings, and Microsoft CEO Satya Nadella said Tuesday that Microsoft-Decision-1.1 was rolling out widely on Azure.

Startups focus on customer data

UK-based Sqwish announced a $3 million pre-seed round. Cofounders Ushnish Sengupta and Federica Freddi met as Cambridge graduate students and had been developing small language models before adapting their architectures for decision-making.

Sqwish says its post-training tools let customers use their own data to improve a model’s judgement. Sengupta told TechCrunch that customers can continue uploading production data so models improve on tasks important to them.

Testing beyond benchmark scores

TypeSafe CEO Diogo Almedia said carefully calibrating Jev’s judgement with statistically adjusted synthetic data was the hardest part of building it. Sqwish also uses synthetic data, while offering customer-specific tuning.

Sqwish’s kettei fat has performed well on some benchmarks, but developers say testing in a specific setting is the clearest measure. Sengupta described benchmarks as an invitation to try a model, with ongoing tuning as a reason to stay.

Models may work together

Microsoft engineering vice president Achint Srivastava said the company balanced cost, latency, intelligence and calibration in Decision 1.1. He cautioned against overemphasizing benchmarks, saying models that score highly on them can falter in enterprise use.

Freddi expects decision models, which she calls System One intelligence, to work alongside frontier language models, or System Two intelligence. She sees a substantial opportunity for decision models, while saying the area is only a side project for much larger companies.

Key points

  • Decision models received more than one in four requests on Vercel’s AI model router in October.
  • TypeSafe’s Jev family leads visible usage, while Microsoft has launched Decision 1.1 for Azure.
  • Sqwish’s tools let customers use production data to keep improving model judgement.
  • Builders say real-world performance and continued tuning matter beyond benchmark results.
Read the original on TechCrunch ↗

More news

Loading stories…