Kimi K3 Becomes The Strongest Open-Weight Model Ever, Built On Scarce GPUs
Moonshot's 2.8-trillion-parameter model ranks third overall behind Claude and GPT-5.6, with weights due July 27.
- Kimi K3 ranked #2 on the Vals AI Index and #1 in Frontend Code Arena.
- New Kimi Delta Attention and Attention Residuals cut training compute roughly 2.5 times versus Kimi K2.
- Analyst Nathan Lambert says Moonshot's team works with real freedom inside a GPU-limited environment.
Why it matters: Chips were rationed, not ideas: scarcity produced discipline no surplus of GPUs ever taught a rival lab.
Nathan Lambert, Interconnects AI ↗ · Jul 21, 20267/21/26 · ✓ Checked✓ Check