Moonshot 推出 Kimi K3.1 Pro,開放權重且比 K3 低 33%。
此版本擁有 2.8T 總引數、104B 活躍引數(MoE)、1M 上下文長度,支援多模態。
AI 伴隨問答
針對這篇報導有疑問?可直接點選快捷問題或向 AI 追問背景脈絡與技術細節:
moonshot dropped kimi k3.1 pro – open weights, 33% cheaper than k3
the specs:
- 2.8t total / 104b active (moe), 1m context, multimodal
- api: $2 input / $10 output per 1m tokens, $0.20 cached (k3: $3 / $15 / $0.30)
- k3.1 pro highspeed: 2–3x faster, $4 / $20
- live on kimi ai, kimi work, kimi code and the api; open weights by nov 4
what's new, per moonshot:
- full-project coding: whole games in unreal, unity and godot
- parametric 3d modeling in blender and cad
- finance work: hallucinations down 65%
early testers:
- ~100 tokens/s on k3.1 pro
- one kimi code session burned 100% of the 5-hour limit on 14.8m tokens (14.7m of them input) – 3% of the monthly cap
- another user says the $99 max plan hit the 5-hour cap after 45 minutes on one prompt
full article in the comments