Open AI Is No Longer Second-Best. That Changes What You Should Build On.
A Chinese company just released the biggest free-to-download AI model in the world, and it's almost as good as the best paid ones — the main hurdle now isn't quality, it's the hardware you need to run it.
- 2.8 trillion
- total parameters — largest open model ever
- 57 vs 60
- K3 vs. top closed model on intelligence benchmark
- $3 / M tokens
- K3 API cost per million input tokens
- 1.4 TB
- storage needed to run it yourself
For most of the past two years, the best AI models were locked behind subscriptions from a handful of US companies. The gap between open and closed AI has narrowed from six months to three — and this week's release makes that shift impossible to ignore. On July 27, Beijing-based Moonshot AI released the full weights for Kimi K3, making the world's largest publicly downloadable AI model free to use.
Kimi K3 has 2.8 trillion total parameters — a 'mixture of experts' design where many smaller models share the work, so only about 104 billion parameters are active on any single task. On the main public benchmark that ranks AI models, it scores 57, just behind Claude Fable 5 at 60 and GPT-5.6 Sol at 59. It ranks first among all open models, and leads in coding tasks. Its context window — the amount of text it can read at once — is one million tokens, enough to hold five long novels.
The practical catch is the hardware. The downloaded model takes 1.4 terabytes of GPU memory to load in its compressed form, and requires a multi-node server cluster with the latest Nvidia or AMD chips. That means self-hosting this model is an enterprise infrastructure project, not a weekend experiment. The license is permissive for most uses, but businesses earning more than $20 million per year from AI services need a separate deal with Moonshot.
For most teams, the immediate option is the API: three dollars per million input tokens — 40 percent cheaper than GPT-5.6 Sol. That makes K3 a real option for anyone doing heavy coding, long-document analysis, or agent workflows. The market has shifted: open AI is now close enough to the frontier that quality is no longer the deciding factor — what matters now is data control, price, and what your team can operate. If you are currently paying premium prices for coding or long-context tasks, K3 via API is worth testing today.