Media Zone | 2026-07-17

Kimi K3's 2.8T parameter count broke the timeline. The biggest open-weight model ever, arriving as the US debates open-model restrictions.

Today's signal

• Dominant story: Kimi K3 is 2.8T params, biggest open-weight model ever (via Play Store listing).
• Pattern: open weights flipping from "catching up" to "biggest model in existence."
• Counter-signal: 2.8T is total params, almost certainly sparse MoE; active count unknown.
• Quiet area: Kurate unchanged, biomedical-heavy; empty Reddit.

LLMs, agents, safety

Kimi K3 at 2.8T parameters

• Biggest open-weight model ever, confirmed via Moonshot Play Store listing.
• For scale: Grok 4.5 1.5T, DeepSeek V4 1.6T, Opus/GPT-5.5 ~1.5T (believed).
• Almost certainly MoE; serving cost depends on active fraction, not the 2.8T headline.
• Lands as the US debates restricting frontier open weights (Interconnects "6 months to live").

Routing, KV cache, compression, GPU

Harness momentum continues

• Harness Handbook (164 upvotes, week's top) still driving the harness-as-artifact conversation.
• The scale-vs-efficiency debate reframed: Kimi K3 may be a sparse-MoE-economics story.