AI
Decision Models Become AI's Hottest Segment as Customization Emerges as Key Differentiator
A new class of lightweight AI models designed for fast decisions is booming, but startups say tailoring these systems to customer data—not raw benchmarks—will separate winners from commodities.
TechCrunch — AI
Decision models have exploded into one of the fastest-growing parts of the AI ecosystem. Last October, more than a quarter of all requests to Vercel's model router went to these specialized systems rather than traditional large language models—a remarkable share for technology that barely existed weeks prior.
What Decision Models Do#
Unlike general-purpose LLMs, decision models operate under a fundamentally different constraint: they don't generate lengthy text outputs. Instead, they leverage their underlying intelligence to rapidly select from a predefined set of options while estimating how confident they are in that choice. This trade-off makes them exceptionally useful in production environments where response speed and cost matter more than detailed reasoning.
Rapid Market Consolidation and Competition#
TypeSafe's Jev family of models has dominated the landscape since its release, achieving a $7.5 billion valuation while handling over 1.5 billion queries across platforms like OpenRouter. Yet the low barriers to entry—one Amazon engineer reportedly built a decision model in a single weekend—have spawned dozens of competitors.
Tech giants moved in swiftly. Microsoft announced it would immediately roll out Microsoft-Decision-1.1 on Azure, while virtually every major cloud provider developed their own version. Smaller players are also entering the ring: UK startup Sqwish closed a $3 million pre-seed round today behind its "kettei fat" model, founded by Cambridge graduates Ushnish Sengupta and Federica Freddi.
Data and Customization as Competitive Moats#
TypeSafe CEO Diogo Almedia emphasized that the hardest challenge in building Jev wasn't the architecture itself but meticulously calibrating judgment using synthetic data. Sqwish is taking a different approach: rather than merely releasing a finished model, the company is bundling post-training tools that let customers continuously refine the system using their own production data.
"What makes us not a commodity is offering our customers the same power—they can keep uploading their production data, and just as we trained our models, their models can also keep getting better at the tasks that they care about," Sengupta explained.

