Kimi K3 Draws Flak for Excessive Token Use

Kimi K3 is being mocked for overly verbose reasoning and generation. One example claims a single SVG output consumed 33,000 tokens, reinforcing the view that Moonshot still has substantial post-training optimization to do.

2026-07-16 ~ 2026-07-17 · 2 related posts

• Kimi-K3 Tends to Overthink During Inference — scaling01 · 2026-07-16
• A Single SVG Consumed 33k Tokens — scaling01 · 2026-07-17