On Kimi K3: Its Capabilities And Related Discontents
Zvi's AI Roundups · Zvi Mowshowitz · 2026-07-20
Zvi Mowshowitz provides a detailed analysis of Kimi K3, Moonshot AI's 2.8-trillion-parameter open-weight model, arguing it is genuinely impressive but remains four to six months behind closed frontier models and is being significantly overhyped as a 'DeepSeek moment.'
Appears in
Extraction
Topics: chinese-ai-modelsopen-weight-modelsai-benchmarksai-geopoliticsai-safety
Claims
- Kimi K3 is approximately four to six months behind the closed model frontier in aggregate capability, with post-training closer and pre-training farther behind.
- Kimi K3's benchmark scores likely overstate its practical performance because all benchmarks were run at maximum effort settings and the model shows jagged, uneven capabilities.
- Kimi K3 is significantly distilled from Claude, particularly Claude Fable, which explains some but not all of its benchmark gains.
- Open-weight release of Kimi K3 carries non-trivial safety risks, with an estimated 10% chance of consequential regret and roughly 2% chance of a serious mistake.
- Recurring narratives that Chinese AI releases have 'erased America's lead' follow an established pattern from the original DeepSeek moment and are being repeated dishonestly or naively.
Key quotes
Do not get carried away. Do not judge Kimi K3 only its relative strengths. In aggregate it is several months behind the closed model frontier, at least four and my median guess is six.
Consider that Kimi K3's (preliminary unofficial) Epoch Capabilities Index is exactly on the Chinese trend line.
I'd estimate something like a 10% chance we regret letting this happen, and ~2% chance that it was a rather serious mistake.