Kimi K3 went live the night of July 16. No model card, no license, no pricing, no benchmark table. Just a flagship model appearing on kimi.com and a timeline losing it.
What’s confirmable: a Mixture-of-Experts model around 2.5 trillion parameters, 1M-token context, text plus images and audio. A general-purpose chat and coding model, aimed straight at where GPT and Claude make their money.
What testers are claiming
Beats GPT-5.5 on some coding evals. Stronger on spatial reasoning and 3D generation than GPT-5.6 Sol and Claude Fable 5. Slow as hell on hard problems — 35 minutes for one task is the number people keep repeating. Note “some.” Third-party SWE-Bench Verified and Terminal-Bench 2.1 results don’t exist yet. Until they do, this is vibes.
Where to use it
kimi.com, in the browser, today. The API is live but Moonshot hasn’t published per-token pricing. The K2 family shipped open-weight under a modified MIT license, so everyone assumes K3 follows. Assumption, not fact.
Why it matters
Three months after K2.6, at a $20B valuation and $300M ARR, Moonshot shipped 2.5T parameters with no paperwork. HN is calling it another DeepSeek R1 moment (218 points). Premature. But DeepSeek V4-Pro already proved the pattern — reach parity, then price the gap to zero. K3 is the next test.
You Might Also Like
- Kimi k2 6 Beats gpt 5 4 and Claude Opus 4 6 on swe Bench pro
- Gpt 5 6 sol Ultra Hits 91 9 on Terminal Bench 2 1 and its Landing in Codex
- Claude Code Remote Control Just Turned my Phone Into a Coding Terminal and im Weirdly Into it
- Cursor Composer 2 Takes on Anthropic and Openai With a 0 50 m Token Coding Model and the Benchmarks Back it up
- A Single api String Exposed Cursors Secret Composer 2 Runs on Moonshot ais Kimi k2 5
Top AI Product 
Leave a comment