Listen in the Podli app 🎧
Follow your favourite podcasts, listen offline and in the car with CarPlay and Android Auto, and always pick up where you left off. Free to try.
Three companies said the same thing in five days, without coordinating. On Thursday Hugging Face published its State of Open Models report. On Monday Meta gave a capable 30-billion-parameter model away for free. On Saturday Bloomberg reported that Stripe has finalized its acquisition of OpenRouter, a company that builds no models at all, for more than $7 billion.
One consistent verdict: the weights are becoming the cheap part of the AI stack. The strategic question inside a company quietly changed from which model to pick to who controls the switch, and what it costs to change your mind.
In this episode, Stephen Forte covers:
- Hugging Face's own data: Alibaba's Qwen family at just over two billion downloads so far in 2026, with the compressed builds that run on ordinary hardware at 39.6 million downloads a month against Google Gemma's 20.8 million and Meta Llama's 7.5 million.
- The precision point the coverage missed: Qwen is the dominant modern open-model family, not the most-downloaded model on the platform, and the difference matters.
- Of 28,531 compressed conversions of Alibaba's models on the platform, Alibaba published 54. Strangers made the rest, and what that means when independent shops start making parts for your machine.
- Muse Glimmer: Apache 2.0, no gated download, runs offline on one consumer graphics card. And the detail almost everyone skipped: it is a distilled student model, trained on the outputs of Muse Spark, the more capable model Meta keeps closed.
- The honest credit: Meta's letter commits to an independent board empowered to approve model-release safety criteria. Most labs have not put that on paper.
- Why data residency, not ideology, is the honest reason a company runs its own model, told through a Dubai commodities group whose records cannot leave the UAE.
- Stripe paying five times OpenRouter's May valuation in three months, for the layer that makes models swappable, and what a payments company buying the metering seat for intelligence tells you.
- The number to stop trusting: the cumulative AI download count. Four circulating totals, at least three methodologies, and why the only usable figures publish their definitions.
A note on attribution: the Stripe acquisition is reported by Bloomberg; Stripe declined to comment. OpenRouter's user and model counts are the company's own May figures.
Sources:
- Hugging Face, "State of Open Models: Summer 2026," 2026-08-14. Qwen download totals (2,045 million in 2026 across repositories with declared parameter counts), quantized monthly downloads (39.6M vs Gemma 20.8M and Llama 7.5M), 151,448 Qwen derivatives, 28,531 conversions of which 54 official.
- Meta AI Research, "Introducing Muse Glimmer," 2026-08-10. 30B parameters, Apache 2.0, offline on a single consumer graphics card, distilled from Muse Spark.
- Meta, "The Future is for Everyone," 2026-08-10. The governance commitment and the concentrated-power argument.
- Bloomberg, "Stripe Finalizes Deal to Acquire AI Startup OpenRouter for Over $7 Billion," 2026-08-16, with TechCrunch corroboration and Alex Atallah's May description of OpenRouter as "the equivalent of Stripe for AI."
- Previous episode referenced: s1e132, "Software You Did Not Buy," 2026-08-17.
The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.
