What Happened
Moonshot AI published Kimi K3 on July 26: a 2.8-trillion-parameter mixture-of-experts model, the largest open-weight release to date. Early reviewers place it at frontier level for agentic coding, with a 91.2 percent BrowseComp success rate using a single agent.
Why It Matters
Open weights do not mean open operation. K3 is so large that practical deployment requires rented cloud infrastructure, not on-prem hardware. That shifts the value question from 'can we run it ourselves?' to 'who runs it, and what guarantees do we have about their access?'
Practical Choices
- API access: Use K3 through compliant cloud endpoints with data-processing agreements.
- Smaller open models: For most internal tasks, a 7B–34B model on dedicated infra wins on cost, latency, and control.
- Eval parity: Verify that K3’s benchmark strength transfers to your domain before migrating workloads.
What To Watch
China’s open-weight releases are now arriving at frontier cadence. Regulators in the EU and US are mapping these models against national-security review frameworks. Expect access and hosting restrictions to tighten before year end.