An Open-Weight Model Just Beat the Closed Giants on Code — Then Opus 5 Landed

The floor is the number that moved

In three weeks this month, five serious models shipped: GPT-5.6 Sol, Inkling, Kimi K3, Qwen 3.8 Max, and Anthropic’s Opus 5. The headline everyone reached for was that the gap between open and closed AI had collapsed. That’s close to true, and it’s the wrong thing to focus on.

The closed models still hold the ceiling — the single most capable systems, rented by the token. What changed is the floor: the most capable model you can actually own, download, and run. Kimi K3, at 2.8 trillion parameters and open-weight, now lands fourth on one widely used intelligence index, level with the previous Opus and GPT-5.5. Fourth in the world, on weights you can keep. A year ago that floor sat far lower. The floor is chasing the ceiling, and it’s close now in a way it never has been.

Where the reading breaks

The same openness makes the month easy to misread, and it did. Kimi K3 took “number one” — on one coding arena, not overall; it’s fourth on the broad index. Moonshot called it the first open model in the three-trillion-parameter class; the real number is 2.8, rounded up. Alibaba called Qwen 3.8 Max second only to Fable 5, with no published benchmark table behind the claim. Even Opus 5 rewards a careful read: it costs the same as Opus 4.8, posts real gains on hard science and coding, but adds no new category of capability — and the effort dial widely called new actually shipped with Opus 4.5. None of these are lies. All of them are the difference between reading a release and being sold one.

And “open” is doing quiet work. Open weights is not open source, and neither one means runnable. Kimi K3 needs something like a terabyte and a half of memory even compressed; Inkling needs roughly half a terabyte. You own them only if you own a datacenter. Everyone else rents all five through an API. This cycle, “open” means open to a datacenter, not open to your desk.

What actually changed

I spent 25 years in cybersecurity before I started covering this. The number that ever mattered was never how good the best tool was. It was how fast the best tool became the average one. Every time that distance collapsed — exploit kits, ransomware-as-a-service, off-the-shelf tooling — the threat model changed, not because the ceiling rose but because the floor did. That is what open weights do to AI capability. They raise the floor, for everyone, at once. The capability is real. The reading discipline is what most people are still missing.