Kimi K4 Leak Flags Fewer Active Params

The rumor is interesting architecturally but weak as product news; Kimi K3’s real advantage is already aggressive sparsity, so a lower activation count would need to deliver clear gains in cost, latency, or quality.

• –Kimi K3 reportedly activates 104B parameters from a 2.8T-parameter Mixture-of-Experts model.
• –Fewer active parameters could improve inference economics and support longer-context workloads.
• –It could also represent a narrower or differently positioned successor rather than a straightforward upgrade.
• –Developers should treat Kimi K3 as the only shippable baseline until Moonshot publishes official specifications.
• –Licensing, weights, context length, and API availability will matter more than parameter counts alone.