Member-only story
A 295-billion-parameter Chinese open-weights model just dropped at roughly six cents per million input tokens, beat GLM-5.1 in a blind expert review of real work tasks, and is live on OpenRouter right now. And not as a research preview this time. As a release.
That is the part that surprised me, and the part that should change your procurement meeting on Tuesday if your team is the kind of team that runs one model locally behind the firewall. Tencent dropped Hy3 today, the production follow-up to the Hy3 preview it pushed at the end of April, and the combination of price, performance, and license is sharp enough that I am writing this in the same morning the email landed.
The headline numbers, fast: 295 billion total parameters with only 21 billion active at inference, a 3.8-billion-parameter Multi-Token Prediction head sitting on top for speculative decoding, 256K context, 192 experts with top-8 routing, Apache 2.0 license, and day-one availability on OpenRouter for roughly six cents per million input tokens (and you can run it free for the next two weeks while the provider absorbs the cost of onboarding you to its inference farm).
Before anyone in the comments loses their mind: yes, Artificial Analysis has not yet indexed this model. The…