The open-weight arms race hits trillion-parameter scale
Moonshot AI's Kimi K3 (2.8T parameters, open-weights, dropping on Hugging Face by July 27) and Alibaba's Qwen 3.8 (2.4T, previewing now) both landed today, and the HN discussion read like a live scoreboard. The Kimi K3 thread spawned calls for up-to-date LLM cost calculators, skepticism about Chinese models over-thinking simple tasks, and dark-horse jokes about Anthropic open-sourcing Opus just to troll everyone.
The pattern here: every week the gap between "frontier closed model" and "best open-weight model" narrows. Qwen 3.8's announcement was directly triggered by Kimi K3, which itself was a response to earlier moves. This is a tit-for-tat escalation with real consequences for anyone building on proprietary APIs.
There's a counterpoint worth noting: several commenters said Qwen 3.7 Pro was frustrating in daily use, stuck in loops, poor at debugging, and that DeepSeek V4 Pro was night-and-day better. Parameter count is not the same as usability. The race to announce is outpacing the race to actually work.
So what?
If you're building a product on a closed API today, the open-weight alternatives are getting close enough to justify a serious re-evaluation every 60 days. The cost and control advantages of self-hosting are growing faster than most founders are tracking. Don't assume API dependency is a fixed cost of doing business.