Kimi K3 is, at the time of publication, the most powerful open model available. Moonshot AI had announced the weights would be released on July 27 — and they were, right on schedule.
It's a mixture-of-experts model: think of a team of specialists working together, where only a portion is activated at a time. There are 2.8 trillion parameters in total, but only 104 billion are active per use.
Why this matters
- Meeting a deadline on a public launch is a sign of a reliable lab — and it's rare.
- Vision comes built into the model, not as an accessory bolted on later.
- It brings two proprietary architecture techniques: Kimi Delta Attention and attention residuals.
1.Read the spec sheet without getting lost
2 min- 2.8 trillion parameters in total — the size of the whole team.
- 104 billion active — how many actually work on each call. That number is what governs speed.
- Kimi Delta Attention and attention residuals — architecture techniques developed by Moonshot itself.
- MoonViT — the encoder that provides vision capability.
The distinction between total and active is the most confusing one. Total defines how much memory you need; active defines how much each response costs. A 2.8 T model with 104 B active is expensive to store and relatively cheap to use.
Spec sheet
- License
- Aberto, pesos no Hugging Face
- Parameters
- 2,8 T totais / 104 B ativos
- Architecture
- Mistura de especialistas + Kimi Delta Attention
Continue in the full microcourse
You've read the opening of 3 classes
The microcourse covers the complete step-by-step, the selection criteria, where the tool fails, who it's really for — and, in the Expert version, the official address to start today.
- The file size is the obstacle2 min
- How this affects you even without running it2 min
How we verified
We track releases straight from primary sources, transcribe what's demonstrated, check every name and number against the manufacturer's official documentation, and rewrite it in Portuguese — with what the tool no do it together, which is the part the ad leaves out.
