American Open-Source Labs Think They Can Beat China’s Best AI Startups

Today on Forbes, American open source labs think they can beat China's best AI startups. Back in October 2024, Poolside was an early AI star. Co-founded by former GitHub CTO Jason Warner, the startup had raised $500 million on a $3 billion valuation to build coding agents for governments and large companies. But over the next 18 months, Poolside largely disappeared from view, while OpenAI and Anthropic ballooned to nearly trillion-dollar valuations, with a crop of Chinese labs building open source models nipping at their heels. Now Poolside is back with a new model called Laguna that, on public benchmarks, beats its American and Chinese open source competition, with the very notable exception of Chinese lab Moonshot's latest AI model, Kimi K3. Warner, Poolside's co-CEO and co-founder, tells Forbes, As an American company building for the West, we'll be the most capable open model in the West. Globally, in this weight class of the 118-billion-parameter model, we are the leader. Warner claims that the startup spent those 18 months, where it went quiet, building the infrastructure for a, quote, model-building factory that could pump out Laguna and continue with new and more powerful iterations every five weeks. He says, quote, The classic notion of model-building is an artisanal process. We've built an industrial model-building process. He concedes that Laguna has a long way to go to match the capabilities of the top closed models from OpenAI and Anthropic, and the new version of Kimi that was dropped by China-based lab Moonshot on July 16. But its models are significantly smaller. OpenAI and Anthropic have not disclosed the size of their latest models, but Moonshot said that Kimi K3 was a 2.8 trillion-parameter model, a whopping 20 times the size of Poolside's new release. Warner says that Poolside plans to build bigger models that could, in time, compete with Frontier Labs. The startup also launched a smaller series of models earlier this year without much fanfare. Warner said it had shared model weights from the get-go with its small, undisclosed number of government and corporate clients, but since April began sharing them publicly. Poolside says it's based in San Francisco, but it has more of an international flavor. Co-founder Warner lives between Canada and San Francisco, while co-founder ISOCant is based in Portugal, with the majority of its team working remotely. The company in the past was reported to be headquartered out of Paris, France, where its coders would assemble every few months to regroup. Silicon Valley is in a rush to build homegrown alternatives to Chinese AI models. These are often billed as open source, but Chinese startups like DeepSeek don't publish their code or training data. Instead, they release what are called model weights, which allow models to be downloaded, run locally, and fine-tuned by developers. That's proven irresistible to startups and increasingly large companies that have chafed at the control OpenAI and Anthropic hold over their own closed models, and the mounting cost of using AI tools like ChatGPT and Cloud. Last week, former OpenAI CTO Mira Murati's Thinking Machines Lab also released its first AI model, an open weight model called Inkling. In a refreshing departure from the AI industry's relentless hype machine, the lab said the model is, quote, not the strongest overall model available today, open or closed, but that it was meant to be flexible and easily customizable. OpenAI, Google, and even NVIDIA are also in the open weight game. Meadow once had one of the leading open models with Llama, but switched its focus to a new closed model called Spark last year. Few have come close to matching the capabilities of the top Chinese models, the most dominant of which is now from Moonshot. Its Kimi K3 outranked OpenAI's top model and Anthropic's Fable 5 at some coding tasks, according to AI leaderboard Arena. It's still behind on some measures, but charges around two-thirds less than Anthropic and a third less than OpenAI for the same task. It is so capable and cheap that its release on July 17th spooked U.S. markets. Over the last month, LLM Marketplace OpenRouter reported that models from Chinese corporations like Xiaomi, Tencent, and startups like ZAI were the five most popular. For now, Kimi K3 is only available to run on Moonshot's servers, located in China, making that a tougher call for U.S. companies. But it said it will publish the model weights on July 27th, which would allow companies to download and run it on their own systems. For full coverage, check out Ian Martin's piece on Forbes.com. This is Kieran Meadows from Forbes. Thanks for tuning in.