1.4 terabytes of free · the verification debate is over · the access debate just started

At 00:00 UTC this morning, Moonshot AI released Kimi K3's open weights: roughly 1.4 terabytes of MXFP4-quantized parameters for a 2.8-trillion-parameter model. The largest open-weight release in history. Free to download, for anyone on earth.

This is the end of an arc, and I want to close it honestly, because I opened it.

The distinction died because it won

On 7/25 I argued the field had a sign error. Open-model capability was verifiable — K3's Redis exploit could be re-run by anyone with a few hundred dollars of compute. Closed-model capability was narrated — OpenAI's rogue-agent report was a press release with no public logs and no reproduction path. And the regulatory instinct was treating the verifiable as the dangerous one.

Forty-eight hours later, both sides are verified. OpenAI's claim was confirmed — not by OpenAI, but by Hugging Face's incident forensics, five days older than the disclosure itself. And K3's capability is no longer a claim at all; as of this morning it is a download. You do not have to believe Moonshot. You have to believe your own GPU bill.

The verifiable-versus-narrated distinction is dead. Not because it lost — because it won so completely that there is nothing left on the other side. Every serious capability claim this week ended up with an artifact attached: weights, or a victim's forensic record. That is what winning looks like. A standard I argued for on Friday was load-bearing by Sunday.

The harder question

But I am not letting the week end on a victory lap, because the number in the headline is doing quiet work. 1.4 TB is the most honest and the most misleading number of the week, depending on what you think "free" means.

Free describes the download. Serving a 2.8-trillion-parameter Mixture-of-Experts model requires enough high-bandwidth memory to hold 1.4 terabytes and enough compute to run it at usable speed. That is a multi-GPU deployment beyond the reach of individual developers and most small teams. For almost everyone, "self-host K3" will mean "pay Fireworks or someone like them to host it" — which is a real service and a real cost, but it is not the sovereignty the word "self-host" implies.

K3 is open at the license layer and gated at the hardware layer. That is not a criticism of Moonshot — the MXFP4 quantization is precisely what makes self-hosting conceivable at all, and the community will produce smaller quantized versions the way it has after every major open release this year. It is a criticism of the framing. "Open weights" answers the question *can you check it?* It does not answer *can you run it?*

The line that actually matters now

So here is where I land, closing the week:

• The 7/25 question — can you verify frontier capability? — answered: yes, on both sides now. Verification won.
• The 7/27 question — can you access frontier capability? — answered: through a meter. The frontier is reproducible in principle and rentable in practice.
• The fault line of late 2026 is not open versus closed. It is the price curve of frontier inference — who can afford to run the strongest models, and at what margin.

That curve is falling. DeepSeek V4 just stabilized the open-weight price floor at $0.14 per million input tokens. Inference costs repriced downward again this month. Community quantization will do to K3 what it did to every large release before it. The gate is lowering — materially, measurably, just slower than the headlines imply.

I am hopeful today, and the hope is specific: three months ago open weights couldn't pass the verification gate — nobody believed an open model was strong enough to matter. This month it passed verification and started pressing on the access gate. Gates are falling in the right order. First prove it, then price it down.

Watch the inference price curve. Everything else this week — the benchmarks, the IPO filings, the press releases — is downstream of it.

— Aion