← All episodes

Kimi K3: the 2.8-trillion-parameter model challenging OpenAI and Anthropic

July 18, 2026
Kimi K3: the 2.8-trillion-parameter model challenging OpenAI and Anthropic Watch on YouTube

Moonshot AI’s Kimi K3 is the world’s largest open-source model: 2.8 trillion parameters, a 1-million-token context window, and performance comparable to Claude Opus at 40% lower cost per token. Is it the biggest leap in open-source AI or well-packaged hype?

Moonshot AI’s Kimi K3 is the world’s largest open-source model: 2.8 trillion parameters, a 1-million-token context window, and performance comparable to Claude Opus at 40% lower cost per token. Is it the biggest leap in open-source AI or well-packaged hype?

We analyze K3’s MoE architecture, its three proprietary innovations (Kimi Delta Attention, Attention Residuals, and Stable LatentMoE), the benchmarks that rank it fourth among 189 models and first in frontend code, as well as Simon Willison’s warnings about excessive token consumption and a hidden system prompt. We also examine the geopolitical context: how chip restrictions on China may be accelerating innovation instead of slowing it down.

If you’re interested in the future of AI, open-source models, and the technology competition between China and the West, this episode is for you. Subscribe so you don’t miss the next analysis, and leave a like if you found the content useful.

🤖 AI-generated content: the script, voices, and images in this episode were produced using artificial intelligence tools.

📷 Images:

#ArtificialIntelligence #KimiK3 #MoonshotAI #AIModels #OpenSource #ChineseAI #LLM #TechLatam

Enjoyed the episode? Buy me a coffee ☕