Skip to main content

Kimi K2.5 Bids Farewell: Moonshot AI's First Trillion-Parameter Model Retires

Kimi K2.5 Bids Farewell: Moonshot AI's First Trillion-Parameter Model Retires

It's the end of an era for Moonshot AI. The company announced today via Kimi's official Weibo that Kimi K2.5, its first trillion-parameter multimodal model, will officially retire at the end of this month. Launched and open-sourced in January this year, K2.5 was a trailblazer—the most versatile model in Kimi's history up to that point. It featured a native multimodal architecture that could handle both visual and text inputs, switch between thinking and non-thinking modes, and tackle everything from casual dialogue to complex Agent tasks.

Image

K3 Takes Over, Parameter Scale Rises Again

K2.5 wasn't just a pretty face; it achieved state-of-the-art performance among open-source models on tasks like Agents, code, images, videos, and general intelligence. It was designed as a general-purpose multimodal model, seamlessly transitioning between conversation, complex reasoning, and automation. Its retirement marks a clear shift in Moonshot's focus toward its successor, Kimi K3, which was open-sourced in July.

K3 is a beast: a 2.8 trillion parameter model built using the KDA hybrid linear attention mechanism and attention residual technology. It natively supports visual understanding and boasts a context window of up to 1 million tokens—enough to process entire books in one go. The official statement highlights that K3 is the first open-source model to reach the 2.8 trillion parameter scale, and for the past 12 months, Kimi has held the record for the largest open-source model scale for 9 of those months. That's a pretty impressive track record.

So, what does this mean for developers and AI enthusiasts? If you've been using K2.5, you'll need to migrate to K3 before the end of the month. The good news is that K3 is already available and open-sourced, so you can start experimenting with its enhanced capabilities right away. The retirement of K2.5 is a natural part of the AI lifecycle—models evolve, and newer, more powerful versions take their place. But it's also a moment to appreciate how far we've come in just a few years.

As we say goodbye to K2.5, we welcome K3 with open arms. The future of AI is bright, and Moonshot AI is clearly at the forefront, pushing the boundaries of what's possible. If you're curious about K3's capabilities, there's no better time to dive in. After all, the next big thing is already here.

Key Points

  • Kimi K2.5, Moonshot AI's first trillion-parameter multimodal model, will retire at the end of this month.
  • Launched in January, K2.5 was known for its versatility, supporting vision, text, thinking and non-thinking modes, and Agent tasks.
  • Its successor, Kimi K3, open-sourced in July, features 2.8 trillion parameters, a 1M token context window, and native visual understanding.
  • K3 is the first open-source model to reach the 2.8 trillion parameter scale, continuing Kimi's record for largest open-source models.