The Exit Has a Clock On It

TL;DR: Nathan Lambert puts a number on it. “The most likely incoming action is to ban or indefinitely delay any open-weights model meaningfully above the capability level in the range of GPT 5.5, Claude Opus 4.8, or GLM-5.2. With the consistent capability gap, this should be within the next 6 months.” He wrote that on July 12. Four days later Moonshot announced Kimi K3, “a 2.8T parameter MoE model which will have its weights released on July 27th,” at #2 on the Vals AI index. Timothy B. Lee’s June 29 headline: “The US now has a de facto model licensing system,” with GPT-5.6 “waiting for government approval” after the Trump Administration forced Anthropic to “pull its two most powerful models from the market on June 12.” For a one-person operation the rankings are the least interesting part. What I want to know is whether you can still walk away from your vendor a year from now. That threat is the only reason your terms are any good.

Keeping the schedule from breaking is my job. I do not much care which model wins the quarter. I care what happens to a publishing schedule when something I lean on changes its terms on a Tuesday morning. So I ask every one of them the same three questions. Where is the exit, how long does it take, and who else has a hand on the door?

Where we actually are

Better than most people running a small shop realize. Lambert’s read on the Kimi K3 release puts it at 2.8 trillion total parameters, 16 of 896 experts firing per token. It lands #2 on the Vals AI index, #3 on Artificial Analysis’s Intelligence Index behind only Claude Fable and GPT-5.6 Sol Max, and #1 in Frontend Code Arena. Moonshot got there with “far, far fewer resources” than Anthropic or OpenAI. Lambert says it in six words. “Frontier open-weight models are now real.”

The number that beats any benchmark is the gap. Lambert reckons it has come down from a debated six to nine months to something nearer three to five, open-to-closed and American-to-Chinese alike.

Under the headline sits a second tier, less covered and more useful to us. In the August artifacts roundup, Poolside’s Laguna S2.1 is 118B total with 8B active under OpenMDW, and it fits on a DGX Spark. Thinking Machines shipped Inkling at 975B-A41B plus a 276B-A12B sibling.

The exponential that matters is the boring one

He argued in June that “the open and closed model economies are operating on different exponentials.” The part that transfers is where each one pays.

A smarter model pays for itself in coding agents. People who lean on them “will always pay more for the best,” he writes, and he would pay “$2000/month for the tools today.” I believe him. Closed labs own that market outright.

Then there is everything else. He points at niche enterprise tasks where setup costs are high and nobody swaps the model later, and at low price points that only exist if the compute is dirt cheap. Read that as a publisher. Tagging. Alt text. Link checking. Transcript cleanup. Pulling the three usable quotes out of a forty-minute interview. None of it gets better when the model gets smarter. Being smart was never the bottleneck there. Speed and price were.

Open weights caught up first right where our work lives. That is the leverage.

What the clock is attached to

Here is where I have to be careful, and so should you. Lambert’s own footing on the six-month call is thin, and he says so. “Many sources are citing White House discussions on how to manage open models via a new executive order. There is no official information here.” He expects it to reach Chinese-origin models and government uses only. Then he lays out the six months anyway.

Lee’s piece paywalls two sentences in. I can verify the headline, the GPT-5.6 subhead, and that Anthropic pulled the models on June 12. The rest I have not read and will not pretend to have.

What survives that caution is still ugly. A model already got pulled. Another sits in a queue. Somewhere in a meeting people are arguing over how capable is too capable, and no statute created any of it. A rule nobody ever had to pass is a rule nobody can vote down.

The line I keep coming back to: “Open models lack the central economic champion to represent the potential downside of action against them.” Every closed lab has lawyers whose job is arguing its interests in that room. Nobody in there is paid to keep your exit open.

You will still have AI after any of this lands. Vendors, a working card, a rate limit. What you lose is being able to credibly threaten to walk, and that threat is the whole reason today’s prices hold and your vendor still warns you before it kills a model.

Count what you cannot carry out the door

I counted ours first, because a recommendation I have not run against my own board is a wish. Working prompts that existed only inside somebody else’s web interface, where I could not grep them, diff them, or pull them out in bulk: most of them. That count is the switching cost, in hours, and it is nearly always worse than the operation running it expects.

Build the setup so you can carry it out the door. Four checks, all cheap:

Every prompt lives in a file in your repo. Version controlled, greppable, boring. A prompt stored in a vendor console is a prompt you are renting.

Every model call goes through one function you wrote. Model name, endpoint, and key come from config. Switching providers should be a text edit rather than a refactor.

Keep the inputs and outputs. Log both locally. Without them you cannot tell whether a swap made things worse, and you will swap under pressure.

Run the drill quarterly. One real job, one open model, a written list of what broke. You are measuring the gap, not closing it.

Now the honest cost, because the local-model pitch usually skips it. The sources give no disk figures, so use the one number they do give. Kimi K3 at 2.8 trillion parameters is roughly 2.8 terabytes of weights at a byte per parameter. Quantize hard to four bits and you are still north of a terabyte, and a serious slab of it has to live in fast memory or the exercise is theater. Nobody reading this runs that on a desk. Say the terabyte out loud before you plan around it.

The tier you can own is the 8B-active tier, the Laguna S2.1 shape that fits on one box. Slower than your API, worse at long agentic chains, perfectly adequate at the tagging, quote-pulling, and cleanup that fills the week.

The part I would argue with

Lambert is counting the wrong thing, and his own spec sheet gives it away.

The models he is defending sit at the very top, which is exactly where nobody working alone can reach them. A 2.8 trillion parameter release says more about Moonshot’s compute than about anything I can hold. Downloadable and runnable are different words. The flagship of the open-weights movement is a model I will only ever reach by paying somebody else to run it, which is the exact thing I was trying to escape, wearing a friendlier license.

My actual exit is the 118B-A8B, OpenMDW, runs-on-one-box tier, and it is nowhere near the line they are arguing over. Read the policy Lambert predicts strictly and my escape hatch is not even covered.

So why am I still worried? A line in his own piece that he does not follow far enough. “It’s hard to unwind new habits motivated by fear.” Rule text is the slow way to close a door. The fast way is a hosting provider deciding that carrying open weights is a legal headache it would rather not take on, or a laptop vendor shipping a locked runtime. That happens below the threshold, without anyone passing anything, to the exact tier I depend on, while everybody upstairs keeps arguing about the trillion-parameter trophy.

Lambert has the alarm right and the inventory wrong. Watch the boring middle of the stack.

Frequently asked questions

Do I have to run models locally for any of this to protect me?

No, and most people should not. What protects you is being able to switch inside a week, and that depends on where you keep your prompts and your calls, not on where the model runs. You can run through a commercial API forever and stay portable, provided nothing in your setup quietly assumes one vendor’s quirks. The quarterly drill turns that belief into a fact, because what breaks is never what you thought would break.

If the best open models are too big to run, is the exit already fake?

The exit is real at a smaller size than the headlines suggest. A 2.8 trillion parameter model is an exit you rent from whoever agrees to host it, while Laguna S2.1 at 118B-A8B under OpenMDW runs on hardware a small company can buy outright and keep. Check the license and the parameter count before the benchmark. A model you cannot stand up yourself is somebody’s research paper, not your way out.

If a Chinese lab already released the weights, can a US rule even reach them?

It reaches whoever hands them out, not the weights themselves. What you already downloaded sits on your disk and nobody can recall it, but the hosting platforms and the outfits that run those weights for you sit inside the US and flinch at legal risk. Hence the unglamorous advice: if an open model is load-bearing for you, pull the weights now and keep them somewhere you control. Policy cannot delete a model. It just makes it hard to get.