A Phone Maker and a Food App Now Build Frontier AI

Training data quality determines model quality. Voices delivers custom or pre-built voice datasets—directed by real professional talent, not scraped audio—across any language, accent, or emotional range. 

Choose from 800+ hours of character voice performances, 1,723 hours of QA'd conversational data across 56 emotional states, or 1,000 hours of expressive, multi-language data with full emotion tagging. Every dataset ships with clear licensing, signed consent, and usage rights baked in, so your legal team sleeps easy and your model trains on data it can actually use. 

7 of the world's top 10 AI labs already use Voices. Request free samples and hear the difference professionally directed voice data makes.

Request Voice Data Samples

1.  Hugging Face published the scoreboard, and the download crown is not close.

The platform's summer report shows Alibaba's Qwen family recorded about 2,045 million downloads over the year, roughly 55 times Moonshot's 37 million. The reason is strategy, not just quality: Qwen ships everything from under 1 billion parameters up to the 2.4-trillion-parameter Qwen 3.8 Max, while Moonshot, MiniMax, Xiaomi, and Z.ai publish almost nothing below 70 billion. Cover every size, win every developer.

2.  A phone maker and a food-delivery company both shipped trillion-parameter models.

The genuinely surprising line in the report: Xiaomi and Meituan both cleared a trillion parameters this year, and neither was a household name in open weights twelve months ago. Building large has stopped being a differentiator. When a company known for phones and one known for food delivery can both reach that scale, the moat was never the parameter count.

3.  The size gap between US and Chinese open models is stark.

In almost every month of 2026, the largest open model from a Chinese lab was bigger than anything a US lab released. China's monthly ceiling ran between 754 billion and 2.78 trillion parameters, while US models stayed under 130 billion in five of seven months. American labs are not absent, they are aiming somewhere else, and where they are aiming turns out to matter.

The runnability problem.  Here is the mechanism behind Qwen's 55-to-1 lead. For labs that publish only huge models, a developer's first encounter is with a model too large to run on anything they own. Kimi K3 is the most capable open model available, but at 2.8 trillion parameters it needs a GPU cluster. Downloads measure what people can actually use, which is a different question from what tops a benchmark.

The licence trap.  Capability and licence are separate axes too, and this is where people get caught. Z.ai's GLM-5.2 is a 744-billion-parameter model released under the permissive MIT licence, while Kimi K3 scores higher but ships under a custom licence with a commercial threshold, as we covered when its weights dropped. One analysis calls GLM-5.2 the right pick for most people precisely because of that licence. Read the terms before the benchmarks.

The American position.  US labs are not competing on raw capability per dollar at the top end, and it helps to be honest about that. Their advantage is jurisdiction and licensing clarity, not benchmark superiority, with Thinking Machines' Inkling the strongest US open-weights entry at 41 on the Artificial Analysis index against K3's 57. If data residency or legal certainty is your constraint, that gap may be worth paying for. If it is not, it probably is not.

The takeaway.  Pick along three axes, in this order: can you run it, does its licence allow what you plan to do, and only then how it scores. Most people skip straight to the third and end up with weights they cannot deploy. And note that self-hosting is what actually solves data residency, because weights running on your own hardware do not send data anywhere, which makes the developer's country far less relevant than the headlines suggest. The prompt below walks through all three.

Qwen passed Llama.  On Hugging Face, the Qwen family has overtaken Meta's Llama models in cumulative downloads, and an MIT study found Chinese open-source models have surpassed US ones in total downloads. Zuckerberg's manifesto last week reads a little differently against that scoreboard.

The 12-hour leaderboard.  If you want the live picture rather than an annual report, there is now a most-downloaded leaderboard pulling from the Hugging Face Hub and refreshing every 12 hours. A useful bookmark for anyone choosing weights this quarter.

Reflection's silence.  Despite reported funding above $2 billion including a round led by Nvidia, Reflection AI has published no open weights, model card, or licence as of August. A company to monitor rather than a model to evaluate, which is a useful distinction to keep in mind whenever a lab announces intentions.

Laptop-class still exists.  Lost in the trillion-parameter arms race: Gemma 4 12B runs on a modern laptop. For summarizing, drafting, and classifying, small local models handle a surprising amount of real work with zero API cost and complete privacy.

Choose an open model by what you can run, not what tops the chart

Benchmark tables are the worst way to pick open weights, because the top entries are usually the ones you cannot run and may not be licensed for your project. This prompt forces the decision through the two gates that actually eliminate options first, and only then looks at scores.