Observed arrival · 2026-10-16
Apex-3 trains on one GH200—in public
ApexSI presents language models alongside their weights, training code, logs, and research.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- Researchers and builders studying open language models
- Worth noticing
- The homepage describes Apex-3 as a 21B MoE with 3.36B active per token, pretrained on one GH200.
Field notes
The homepage links to model listings on Hugging Face and training code on GitHub, alongside logs and research. For Apex-3, it gives a 21B mixture-of-experts size, 3.36B active parameters per token, and says pretraining used one GH200. The logs are described as covering loss, throughput, and MFU, while the research includes what failed. These are ApexSI's descriptions in the supplied page evidence.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
One card from the complete issue