One week, two headlines
Early August delivered two Chinese model stories back to back. One lifted the "largest open" bar straight to 2.4 trillion parameters. The other quietly changed its pricing, and Asian-timezone developers felt it in their bills.
You want to self-host a top-tier model — no API, data never leaves your own servers. Two months ago that felt like a luxury only hyperscalers could afford. Now the weights are open. But pause before you download, because what you can pull and what's topping the leaderboards may not be the same artifact.
Alibaba's Qwen: biggest open release, with asterisks
Qwen3.8-Max totals 2.4 trillion parameters with about 95 billion active per token. Released August 3, weights dropped August 12. The catch: the open checkpoint, Qwen3.8-2.4T-A95B, is a reduced build — text-only, 262K context, while multimodal input and the full 1M-token context live only in Alibaba Cloud's hosted version. The license isn't permissive Apache either; it's a custom terms sheet with revenue gates. MaaS or AI-assistant businesses over $50M rolling revenue need separate authorization.
DeepSeek: capable, now metered like electricity
V4 Pro went general on August 13: 1.6 trillion total, 49 billion active, 1M-token context, up to 384K output, MIT-licensed weights you can host freely. It used to be flat at $0.435 / $0.87 per million tokens. On August 16 it switched to peak/off-peak: peak $1.32 / $3.96, off-peak $0.66 / $1.98. Peak windows are 01:00–04:00 and 06:00–10:00 UTC — which lands at 9am–noon and 2–6pm in Beijing. Basically your working hours.
The translation: a team in Singapore or Beijing calling V4 Pro during the day hits peak pricing almost every time, at roughly triple the off-peak rate. A team in San Francisco doing the same work at night rides off-peak. Same model, cost that varies by where you sit.
The bigger picture: China owns the top five
In OpenRouter's weekly call volume, DeepSeek-V4-Flash led with 7.22 trillion tokens. More telling: Chinese models have held all five top spots globally for fourteen straight weeks. This isn't a demo. People are running it at scale.
Which one should you pick
Want cheap, controllable, friction-free self-hosting? DeepSeek's MIT weights are the calm choice. Want the strongest Chinese multimodal and can live with license terms? Qwen's hosted tier fits. But high-volume Asian teams on a budget should remember DeepSeek's peak lands on your shift — push the heavy jobs to night and you save a real chunk.
Why this matters to you
If you're building your own AI product, wary of being throttled by someone else's API, and need data to stay local, this open-weight wave is a genuine opening. Just settle two accounts first: the license gate (Qwen's revenue net) and the hardware bill (95 billion active params won't run on one card). The opportunity is real; so are the traps.
Related reading: Qwen, DeepSeek, Kimi