Kimi Launches K3 With 2.8T Parameters And 1M Context Window
It supports a 1 million token context window.
Entities: Moonshot AI
Moonshot AI says its new Kimi K3 announcement on X introduces a 2.8 trillion-parameter native multimodal model with a 1 million-token context window, plus technical changes it says can speed decoding in long-context use and improve training efficiency. The launch is getting attention in part because ChrisGPT posted a CS:GO × Portal-style prototype on X, saying K3 generated it in three prompts with about 600,000 tokens and roughly $3.24 in API usage.
Combined views
99.2K
6 posts, first seen 5h ago
Reactions from ranked influencers
6 posts
CS: GO × Portal clone - FPS game
Kimi K3 just 3 shotted this CS:GO × Portal clone for me using around 600,000 tokens. $3.24 in API usage. The same token cost would be $10.80 with Fable 5 & $6 with GPT-5.6 Sol. The era of free indie game development is closer than you think anon! https://twitter.com/kimi_moonshot/status/2077830229968683203
Huge congratulations to the Kimi team, this is amazing!
Introducing Kimi K3: Open Frontier Intelligence 🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal 🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts 🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost 🔹 Built for long-horizon agentic coding and self-evolving workflows Kimi K3 is now live on on http://Kimi.com, Kimi Work, Kimi Code, and the Kimi API. Open Weights by July 27, 2026. 🔗 API: http://platform.kimi.ai 🔗 Tech blog: http://kimi.com/blog/kimi-k3
German car industry for the past 20 years: "The Chinese™️ can only copy" American AI labs now: "The Chinese™️ can only distill"
If you still dismiss Kimi as just a “distiller”, you are very wrong and totally miss the point. After reading the Kimi K3 blog, Kimi makes many genuine, hardcore architectural innovations: * KDA hybrid linear attention (for efficient long context scaling) * Attention Residuals (a more efficient memory retrieval) * Stable LatentMoE (only 1.8% experts will be activated each time) * Quantile-balanced routing (infra optimization) 2.8T, it will give the world’s largest open-weight model. A new architecture playground. A co-design among architecture, training, serving and agent. So excited.
You can use K3 as a teacher model (with logits!) to train up smaller dense models. It’s actually feasible even to use it to create a trillion token full pre and post training dataset for even better models.
For people who doubted KDA K3 is a scarily good base, built explicitly to carry itself onward, recursively. I fully expect K3.1 to surpass Sol and Fable, and I have no idea where it tops out. Probably at the point when they have enough SuperPoDs to jump to a 10T. (ie 5-10) https://twitter.com/ZhihuFrontier/status/2078286180320247844
These innovations weren’t distilled: * KDA hybrid linear attention (for efficient long context scaling) * Attention Residuals (a more efficient memory retrieval) * Stable LatentMoE (only 1.8% experts will be activated each time) * Quantile-balanced routing (infra optimization)
If you still dismiss Kimi as just a “distiller”, you are very wrong and totally miss the point. After reading the Kimi K3 blog, Kimi makes many genuine, hardcore architectural innovations: * KDA hybrid linear attention (for efficient long context scaling) * Attention Residuals (a more efficient memory retrieval) * Stable LatentMoE (only 1.8% experts will be activated each time) * Quantile-balanced routing (infra optimization) 2.8T, it will give the world’s largest open-weight model. A new architecture playground. A co-design among architecture, training, serving and agent. So excited.