Kimi's next model could pack 2.5 trillion parameters and 1M context

X account @gpuhell claims that Kimi K3 is planned for a Q3 release with over 2.5 trillion parameters. Internal experiments have tested context lengths far beyond 1 million tokens, but it's unclear if the final public version will offer 1M context. The leaker also says the main bottleneck preventing Kimi from shipping 1M context today isn't technology — it's compute resources.

DeepSeek's V4 series has already made 1M context a public selling point, raising the bar for open-weight models on long context. Kimi's current public version, K2.6, still has a 256K context window. If K3 really brings 1M context to the public, Moonshot will catch up to DeepSeek V4 in both model size and long-context ability.