Kimi K3 ties Opus 4.8 and GPT-5.6 on Agent Arena; open Jul 27
The largest open-weight model yet — 2.8T params, ~50B active — already tops Frontend Code Arena; full weights drop in four days, if your GPUs can hold it.
Copy markdown
A frontier tie, in the open lane
Arena.ai's Agent Arena now ranks Kimi K3 #4 overall — level with Claude Opus 4.8 and GPT-5.6 Sol — after a jump from #23. It's the first open-weight model to reach the closed frontier's tier on agentic tasks.
2.8T params, ~50B active, open on Jul 27
K3 is the largest open-weight model to date: 2.8T total with 16 of 896 experts (~50B) active per token, a 1M-token context, and native vision. It already sits #1 on Frontend Code Arena at 1,679 points, ahead of Claude Fable 5. Full weights are promised July 27.
The self-host asterisk
"Open weights" here is a promise with a date, not a laptop download. Moonshot recommends supernodes of 64+ accelerators to serve K3 at speed; for most builders that means a hosted endpoint, a cloud deploy, or a heavily quantized derivative — not full local inference.
Demand already broke the hosted tier
Moonshot paused new subscriptions after usage surged over 48 hours and strained its compute. If you're wiring K3's API into a product before the weights land, plan around capacity limits now.
What you could ship
Once the weights are out, frontier-class coding and agentic performance becomes something you can own, fine-tune, and serve through any provider — no per-token rent to a single lab. Even a quantized K3 on rented GPUs could back a private coding agent you fully control.