Open models
Tencent posts Hy4’s open weights, and they need about 944 GB of GPU memory
Tencent says its 770B Hy4 reads images, drives a computer and writes game code. The weights are free to download, but running them takes an 8-GPU server.

Open models
Tencent says its 770B Hy4 reads images, drives a computer and writes game code. The weights are free to download, but running them takes an 8-GPU server.

Explainer
Small language models now run on a work laptop. What they need in memory, where vendor benchmarks show them holding up, and where they still fall apart.

Open models
Nous Research's open Arete-27B retunes Qwen3.8-27B. On Nous's own 803-prompt test, propaganda answers fell from 22% to 0.9%. No outside replication yet.

Open models
Mistral says its 1-trillion-parameter model leads open-weight coding. Artificial Analysis ranks it second and measured lower scores on two of Mistral's headline tests.
