Kimi K3 dropped two days ago and my timeline hasn’t stopped talking about it. But the actual numbers you need for a build decision — context window, benchmarks, and real token pricing — are scattered across launch posts and docs. I got tired of reconciling five tabs, so I put them side by side.
One table
Kimi K3 Kimi K2.6 Fable 5 GPT Context 1M tokens 256K (262,144) see vendor see vendor API input $3.00 / 1M $0.95 / 1M premium premium API output $15.00 / 1M $4.00 / 1M premium premium Open weights ✅ ✅ ❌ ❌One pricing detail worth flagging: K3’s cache-hit input drops to $0.30 / 1M. If you run long system prompts or multi-turn agents, that’s where the “cheap” reputation actually comes from — not the base rate.
A few things that stood out
- Front-end/UI is its strongest suit. It hit #1 on the Frontend Code Arena (1679) on launch day, above Fable 5. That’s the most consistent praise I’ve seen in the threads.
- It’s a large sparse MoE. Active params are a fraction of total, which is a big part of why the per-token price is what it is — it’s not “a smaller model,” it’s cheaper inference.
- Self-hosting isn’t realistic for most people. The total parameter count is datacenter territory; the practical routes are quantized community builds or hosted inference (OpenRouter is a couple of bucks to just try it).
Caveats
Where I couldn’t confirm an official number I left it blank instead of guessing. Pricing is per Moonshot’s stated rates — verify against the official pages before you build on it. Benchmarks are one signal, not the whole story; for “does the output actually feel good,” throw a couple of dollars at your own hardest task.
Full table and more fields here (disclosure: I built the page): https://kimi3.online
If you’ve run K3 in anger, I’d love corrections — especially on backend/reasoning vs Sol 5.6.
답글 남기기