Modest models, well-guarded vaults: the finesse layer is the prize

A tricky time to be a banking CEO.

Last Friday, a Chinese startup shaved a percentage point off the S&P 500 with a model launch. Moonshot AI unveiled Kimi K3, an open-source model the company says closes much of the gap with OpenAI's ChatGPT and Anthropic's Claude. On the news, the Nasdaq dropped 1.4% and Taiwan's benchmark index closed down more than 6%.

You might recall DeepSeek triggering similar market ripples in January 2025, although US markets recovered quickly enough.

At this point, it’s evident that signs are pointing to frontier-grade LLM prowess commoditising, with Kimi K3 serving as yet another harbinger for enterprise incumbents. That begs the question, if the smartest model in the room may soon be free, is the LLM space race to global dominance any more than a vanity dash to nowhere-ville?

Three weeks before last Friday's Kimi K3-induced excitement, the AI research unit at the world's largest hedge fund, Bridgewater, working with Mira Murati's Thinking Machines Lab, published results showing that a fine-tuned version of Qwen3-235B, Alibaba’s open-weights base model, outperformed every frontier model it tested on six financial judgment tasks drawn from investors' daily work.