What Are Open Weights? Kimi, GLM, DeepSeek, and the Models You Can Own.

In June, the best coding model on earth was switched off for twenty days. Every customer, everywhere, at once. Not because Anthropic wanted to. Because a US export-control directive said so.

That is what renting intelligence means. The vendor's obligations sit between you and the model, and when their situation changes, so does your stack.

Open weights are the other option, and in 2026 they stopped being the compromise option. This post is the foundation for a series on them: what they actually are, which models you can genuinely download today, and what changed this summer.

What "open weights" actually means

A model is, physically, a very large file of numbers. The weights. Training produces them; inference runs them. Whoever holds that file can run the model, on their hardware, forever.

There are three tiers of access, and the differences are the whole subject:

API access. You send tokens to someone else's computer and pay per million. You never touch the weights. You cannot inspect the model, cannot fine-tune it beyond what the vendor offers, and your access exists at the pleasure of the vendor, their government, and their uptime. This is renting.

Open weights. The weights file is published, usually with a technical report, under a licence such as MIT or Apache 2.0. You can download it, run it on your own machines inside your own network, fine-tune it on your own data, and nobody can switch it off. What you do not get is the recipe: the training data and full training code stay private. You own the artifact, not the process that made it.

Open source. Weights plus the recipe: training code, and at least a description of the data. Rare at the frontier. Most of what gets called "open source AI" is actually open weights, and the distinction matters because you can verify and reproduce an open-source model but only use an open-weight one.

The models you can actually download

The roster changed completely between June and August 2026. One column matters more than the specs: whether the weights are actually published, or only promised. Vendors have learned that announcing openness earns the goodwill before shipping does.

Notice the pattern in the footnotes. Meta opened a 30B distillation while keeping the flagship on the API. DeepSeek opened Flash while V4-Pro stays closed. The open model is increasingly the ambassador for a closed one, which is a real strategy and still a real gift: an Apache 2.0 model that runs on one consumer GPU is a different category of thing from a rate-limited endpoint.

Why this matters to a company that just wants working AI

Four capabilities arrive with the weights file, and none of them are available for rent.

It runs inside your walls. On your infrastructure, inside your network boundary, next to your data. For anyone with sovereignty constraints, procurement data, health records, legal work, this is not a preference. It is the qualifying condition.

It can learn your work. You can post-train an open model on your own workflows and traces. A rented model knows your prompt; an owned one can absorb your patterns.

It prices per workload. Routine volume runs on a cheap open model, frontier judgment goes to the expensive API. Coinbase halved its AI bill this way, defaulting to GLM 5.2 and Kimi 2.7 with smart routing and caching, and Armstrong's point was that it took no usage caps at all.

It cannot be switched off. Twenty days of Fable 5 darkness made this one concrete. A weights file inside your network has no kill switch in Washington, and none in San Francisco either.

Satya Nadella compressed all four into a single test worth pinning above any AI architecture decision:

If swapping models would cost you everything you have built, you do not own your AI capability. You are renting it, and the rent is not the money.

What changed this summer

Three things moved between June and August, and together they are why this series exists.

The gap closed. In June, GLM-5.2 became the first open model to top a major design leaderboard while Fable 5 was suspended. In July, Kimi K3 debuted ahead of a US flagship for the first time ever, and Nvidia's open Nemotron 3 Ultra scored gold-medal level on the International Mathematical Olympiad, 30 of 42 as graded by the IMO team, a first for an open model. Ethan Mollick's note: Kimi K3 and GLM-5.2 would likely also qualify. The often-quoted version: the open-to-frontier gap went from a year, to six months, to six days, inside twelve months. Chamath's sharper version: the capability gap closed much faster than the pricing gap, and the pricing gap remains enormous.

The industry chose a side, almost. On July 24 the letter "Open Weights and American AI Leadership" published with 25 signatories including Nvidia, Microsoft, Meta, Mistral, Palantir, IBM and Hugging Face, urging Washington not to restrict open-weight models before risks are demonstrated. It doubled to 50 signatories within a day as OpenAI and Google joined. Anthropic and Amazon never signed. That letter, and what the signing order reveals, gets its own post in this series.

America shipped, partly. Meta's Muse Glimmer weights in August are the first US follow-through at meaningful quality since the letter. A 30B local model is not the frontier, and the flagship stayed closed. It is still the most consequential American weights release of the summer.

The honest limits

The taste gap is real. On code, open models are near parity: Kimi K3 scores 19/20 on Harbor Town's coding test against Opus 5's 20/20. On design, it scores 14/20 against 20/20. Ethan Mollick's version: the top closed model still leads on judgment in ways benchmarks do not capture. Open weights closed the gap on price and on code. Not yet on taste.

The frontier keeps moving. The six-days number describes one moment, not a settled state. The closed labs have not stopped, and a lot of money is betting they will not.

You trade one dependency for another. Most frontier open weights are Chinese: Moonshot, Z.ai, DeepSeek, Alibaba. The FT spent July asking what the risk of building on them actually is, and it is not a silly question. The honest posture is not "open is safe," it is that a weights file you hold cannot be remotely revoked, whatever its origin, and you should still know its origin.

Open does not mean vetted. The same capabilities that got Fable 5 suspended ship in open models with no classifier in front of them. Owning the model means owning that, too.

The position to take

Not "open is always good enough." Model-agnostic architecture: your context, your skills, your workflows, and your evals built as the layer you own, with the model underneath swappable, rented where judgment matters most, owned where sovereignty or volume demands it. Nadella's test, applied weekly.

The weights are an option now. The loop was always the asset.

Own the option.