K3 claims 6x cheaper throughput at 1M context

K3's Kimi Delta Attention architecture claims up to 6x cheaper throughput at a 1 million token context window, per AINews's roundup on Latent Space.

That number matters more than where K3 lands on any leaderboard. Long-context agent loops call a model over and over inside the same task. A throughput multiple compounds across a run in a way a single benchmark score doesn't. A 6x cut at 1M context is the difference between a long-context agent workflow that's affordable and one that blows a budget.

AINews also has K3 sitting third on DeepSWE, and calls it the first open-weights model to reach frontier level on that benchmark. Artificial Analysis separately scores K3 at 64% on DeepSWE and 84% on Terminal-Bench v2, per the same rundown.

AINews doesn't say what baseline the 6x figure is measured against, or what K3 costs per token in dollar terms. That leaves the throughput claim directional until someone runs it under real load.

K3 is a bet that throughput economics beat leaderboard position for agent-heavy workloads. If the 6x figure holds up outside AINews's report, it undercuts frontier models priced by the token for anything that runs long.