- Dotika
- Posts
- Kimi K2 Thinking leads open source AI
Kimi K2 Thinking leads open source AI
ALSO : Meta brings AI video feed Vibes to Europe
Geoffrey NICHIL
November 07, 2025
News of the day
1. Moonshot AI's Kimi K2 Thinking, an open-source model, now leads key benchmarks in reasoning and coding, surpassing GPT-5 and Claude Sonnet 4.5. → Read more
2. Meta's AI video feed, Vibes, is now available in Europe via the Meta AI app, enabling users to create AI-generated short videos from text prompts. → Read more
3. OpenAI CEO Sam Altman anticipates $20 billion in annual revenue by year-end, fueling massive data center expansion without government aid. → Read more
4. Google plans a secret AI data center on Christmas Island as part of a military cloud deal, strategically positioned for naval monitoring.→ Read more
Our take
Hi Dotikers!
Here’s the week’s turning point: an open model, Kimi K2 Thinking from Moonshot AI, has taken the lead in reasoning and agent benchmarks. This isn’t just another large model; it’s an agent that plans, calls tools, persists over time, and puts real pressure on closed offerings. When an open-source model holds its own on evaluations closest to real-world performance, the economic equation shifts for good.
On the technical side, K2 Thinking is built on a Mixture-of-Experts architecture with 1 trillion parameters, about 32 billion active during inference, native INT4 quantization, and a 256k context window. (Sorry, getting a bit geeky 🙂) The license is a modified MIT that remains permissive, with only a simple attribution requirement once usage exceeds 100 million monthly users or $20 million in monthly revenue. If your CIO can read the license without calling legal, you know the wind is changing.
Why does this matter now? Just days ago, MiniMax M2 was being praised for its agentic performance and efficiency. K2 Thinking pushes things further and confirms a trend that began with DeepSeek R1: the gap between closed and open models is closing fast ; sometimes even flipping. For businesses, the center of gravity is shifting. The key question is no longer which model to buy, but which data to connect, which tools to orchestrate, and how to govern the whole system.
The real test still lies ahead ; in production: robustness, security, source traceability, latency under load. But at this point, paying the rent for a closed API by default is getting hard to justify when an open model matches or exceeds it on realistic tasks. Differentiation will now come from orchestration, integration, and governance.
G.
Meme of the day
Reply