Poolside open-source AI vs Kimi, DeepSeek, Qwen: analysis

Open-source AI used to be the respectable side quest, the thing you did after the API money showed up wearing a Patagonia vest. Now it is the main arena. Forbes reports that Chinese models including Kimi, DeepSeek and Qwen are dominant AI tools, while American upstarts such as Mira Murati’s Thinking Machines, Reflection AI and Jason Warner’s Poolside are trying to answer with their own open-source push. Poolside is the spicy one here, partly because it went quiet after arriving with fog machine energy. According to Forbes, the company was cofounded by former GitHub CTO Jason Warner, raised $500 million at a $3 billion valuation back in October 2024, and set out to build coding agents for governments and large companies. Now it is back with Laguna, a new model that Forbes says beats American and Chinese open-source competition on public benchmarks, except for Moonshot’s Kimi K3. In other words, the American comeback story has a footnote, because AI benchmarks love nothing more than humiliating clean narratives.

Forbes: Poolside returns with Laguna Forbes frames Poolside’s return

as part of a wider American effort to challenge Chinese open-source leaders, not as another lonely model launch drifting through the benchmark aquarium. That matters because the company is not merely trying to be a cheaper chatbot with a flag pin. Forbes says Poolside was built around coding agents for governments and large companies, which is a much narrower and more useful target than the usual enterprise pitch of helping everyone write emails slightly faster (civilization saved, please clap). The Laguna claim is also strategically pointed. Forbes reports that Laguna beats its American and Chinese open-source competition on public benchmarks, with the exception of Moonshot’s Kimi K3. That caveat is doing real work: it says U.S. open labs can be competitive, but also that Chinese labs are not politely waiting for Silicon Valley to finish its keynote deck. If Poolside wants to win, it needs more than a scoreboard. It needs distribution, trust, deployment flexibility and a reason for builders to choose it when Kimi, DeepSeek and Qwen are already in the toolbelt.

MIT Technology Review: China made openness a distribution engine MIT

Technology Review explains why Chinese labs gained so much developer mindshare: they made their best models free as downloadable open-weight packages. Developers can adapt those models and run them on their own hardware without negotiating a commercial relationship with a U.S. API gatekeeper, according to the publication. That is not just generosity. It is product distribution wearing a lab coat. This is the part American labs have to understand without turning it into a patriotic bake sale. Open weights reduce friction, especially for teams that need local deployment, customization or control over data flows. A closed API can be elegant, but sometimes it is like renting a kitchen where the landlord charges per onion. For coding agents, the ability to run, modify and integrate deeply into enterprise systems could become a serious wedge, especially if buyers do not want their software supply chain mediated entirely through a few frontier model providers.

South China Morning Post and CNN: the lead is real, not magic The South China

Morning Post reported on a Stanford University DigiChina Project report that found Chinese open-source AI models may have caught up or "even pulled ahead" of U.S. counterparts in capabilities and adoption. The same report said Chinese-made open-weight models are "unavoidable" in the global competitive AI landscape and urged U.S. actors not to avoid "selective engagement" with Chinese firms, academics and policymakers. That is a remarkably sane sentence for AI geopolitics, a field that often sounds like someone fed a defense memo into a smoke alarm. CNN adds useful cold water. At a Beijing meeting in January, Justin Lin, technical lead for Alibaba’s Qwen models, estimated the chance of a Chinese AI firm overtaking U.S. frontrunners in the next three to five years at "Below 20 percent," adding that "20 percent is already very optimistic." So yes, China’s open models are strong and widely adopted. No, that does not mean every U.S. lab should immediately replace strategy with panic confetti.

Bruegel: the battle moved up

the stack Bruegel’s Alicia García-Herrero and Bertin Martens argue that the U.S. and China rivalry is moving beyond chips alone, with China challenging U.S. leadership in both AI hardware and software. Chips are Theo’s swamp, and may he emerge from it with wafer-level packaging diagrams and a thousand-yard stare. For the rest of us, the key point is that model access, software ecosystems and deployment channels are now strategic infrastructure too. That is why Poolside’s positioning is worth watching. Open-source models are not automatically better, just as leaving your garage door open does not make your house a community center. But openness can be a distribution tactic, a trust signal and a practical requirement for builders who need control. If Poolside can pair Laguna with useful coding-agent workflows for large organizations, it may have a sharper lane than trying to out-chat the giants. For readers building with AI, the takeaway is simple: stop treating open models as the discount aisle. Evaluate them as deployable components with tradeoffs in performance, control, cost and integration. Watch whether Poolside, Thinking Machines and Reflection AI can turn American open weights into durable developer ecosystems, and whether Kimi, DeepSeek and Qwen keep setting the tempo. The AI race may not be won by the biggest model, but by the one developers can actually take home and make weird.