In October, the AI developer community erupted in celebration when two of Silicon Valley’s hottest coding tools, Cursor and Windsurf, unveiled their “first self-developed models.” But the excitement didn’t last long—sharp-eyed users quickly noticed something odd. These so-called “in-house” models began generating Chinese text during reasoning, and some even admitted, when prompted, that they were originally developed by China-based Zhipu AI. The revelation sparked online jokes: “They open-source, then claim it’s their own.” At the time, such borrowing felt like a clumsy attempt to hide—like repainting a borrowed car to pass as new. But now, that secrecy seems gone. In a bold move, San Francisco-based startup Deep Cogito announced its latest flagship model, Cogito v2.1 671B. CEO Drishan Arora declared on X: “Today, we release the best open-source large language model built by an American company.” To back up the claim, Arora shared impressive benchmark results: the model nearly matches GPT-5 on GPQA Diamond, outperforms Claude Sonnet 4.5 on MMLU across multiple languages, and surpasses Meta’s Llama series in math and code tasks. The charts looked promising—almost enough to believe in a U.S. comeback in open-source AI. Yet, for those familiar with the AI landscape, the number 671B rang a bell. It’s precisely the parameter count of DeepSeek-V3. Soon after, users discovered a line in the model’s Hugging Face configuration file: “base_model: deepseek-ai/DeepSeek-V3-Base.” Unlike earlier cases where companies denied or obscured their reliance on foreign models, Deep Cogito made no attempt to hide it. Arora openly acknowledged the model was forked from DeepSeek-V3-Base. He argued that pre-training has become a commoditized process—like electricity—while the real challenge lies in post-training: refining a base model to reach frontier-level intelligence. “Truly competitive open-source models are rare,” Arora noted. “Outside Meta, there are almost no viable options in the U.S.” Given DeepSeek’s strong ecosystem for efficient inference, it was a logical choice. So what exactly did Deep Cogito build? From the start, the company—founded by a former DeepMind product manager and a Google senior engineer—never aimed to pre-train from scratch. They recognized that most base models today perform similarly. The real competition is now in post-training. Deep Cogito’s core innovation lies in its “frontier post-training stack.” They took DeepSeek’s base model and applied a proprietary reinforcement learning framework combined with Iterated Distillation and Amplification (IDA). Using hundreds of GPU nodes in a large-scale distributed setup, they retrained the model with a focus on efficiency and reasoning depth. Compared to its predecessor, Cogito v2.1 delivers significant gains in inference efficiency. On complex reasoning tasks, it uses just 4,894 tokens on average—less than half of Google’s Gemini 2.5 Pro, which consumed 9,178 tokens. This efficiency stems from “process supervision,” a technique that trains models to think more intuitively. Instead of lengthy chain-of-thought reasoning, Cogito v2.1 learns to find correct answers faster, minimizing unnecessary computation—a hallmark of the IDA approach. On the MATH-500 benchmark, Cogito v2.1 scored 98.57%, narrowly edging out its “teacher” model, DeepSeek-V3-2 (97.87%), and far surpassing Llama 4 Scout. It also performed well on SWE-Bench Verified, a code repair challenge. Objectively, Cogito v2.1 is a strong model—particularly in reasoning and efficiency. Deep Cogito’s technical work in post-training is commendable. They’ve demonstrated that with the right methods, a model can be pushed to elite levels, even if it starts from someone else’s foundation. What’s more, the company has been transparent: the base model is clearly cited, and no effort was made to conceal its origins. Still, calling a model built on DeepSeek’s architecture “the best open-source model made in the U.S.” stretches credibility. The core design and parameters are Chinese. What Deep Cogito contributed is refinement—not invention. Even more troubling is the subtle ideological shaping embedded in the model’s behavior—something that undermines the spirit of open-source collaboration. True innovation thrives on transparency and shared progress, not national branding. Deep Cogito’s move reflects a broader trend: China’s open-source models are now the de facto foundation for many global AI startups. Pre-training a 671B model from scratch costs tens of millions. Building on an existing, high-quality model is far more practical—cost-effective and fast. This pragmatic choice is not wrong. But the desire to claim “American-made” excellence while relying on Chinese technology reveals a deeper tension. It reflects the anxiety of Silicon Valley’s once-dominant players facing a new reality: innovation no longer flows only from the West. The real test of technological confidence isn’t in national narratives, but in honesty, contribution, and collaboration. Open-source thrives not when models are rebranded, but when they’re shared, improved, and advanced together—across borders.