Korva is a scroll of the internet’s best ideas.

Where America ranks

Arena is the platform that ranks AI models by what real users prefer. The American model on top of its open-source list is Inkling, from Thinking Machines, a lab that is about a year and a half old. Inkling came out of a team restructuring roughly six months ago, after which the company restarted its work, so the top American spot took about half a year.

What buyers ask

The gap matters to buyers. Large enterprises are wary of both the frontier labs and Chinese open source. Arena was pitching its products to a Fortune 50 company when the company asked whether anything in the stack was built on Qwen, a Chinese model, and whether Arena could switch to an American one. Arena could.

Other American models

Among the other American open models, Google's Gemma shows up on Arena's Pareto curves of performance against cost, which marks it as efficient.

Why do Chinese labs release their models openly?

Releasing weights spreads a lab's model quickly, since anyone can download it and build on it. Developers who build on one model tend to keep using that lab's later releases, which gives the lab influence beyond its own customers. Meta's Llama line followed similar logic earlier. Alibaba released many of its Qwen models under the Apache 2.0 license, and DeepSeek released its R1 model under the MIT license, and both licenses allow commercial use.

How does a Pareto curve compare models?

A Pareto frontier marks every model for which no other model is both better and cheaper. Plotting quality against cost, the models on the frontier are the efficient choices, and the rest sit below it. A model can lead on quality yet appear nowhere on the frontier if a cheaper model matches it. The concept is named for the Italian economist Vilfredo Pareto, who studied efficient allocations around 1900.

How do big companies vet a model's origin?

Large enterprises run vendor risk reviews that ask where software comes from and what it depends on. Executive Order 14028 on cybersecurity, signed in May 2021, pushed the software bill of materials, a list of every component inside a product, into mainstream procurement. A model built on a particular base model is one more component to list. Such reviews usually happen before a contract is signed.

More from 20VC with Harry Stebbings

Business & StartupsNeo labs need $4 billion in revenueA $10 billion neo lab has to become a $100 billion business to pay investors 10x, which takes at least $4 billion in revenue within two or three years.
Enjoyed it?