We Bet the Company on Open Weights. Here's What's Happening to Them.

Our product has a promise in it that I can't keep by myself.

Vinci Forte serves the best open-weight model available.

Not a model we picked and froze. Whatever is best right now, verified by our own evals, at cost plus 15%.

When something better ships, Forte becomes that thing and you don't migrate anything.

That promise has a dependency I don't control.

It assumes there will keep being a best open-weight model.

For the last two years that was such a safe assumption nobody wrote it down.

In the last four months it stopped being safe.

And almost nobody is talking about the part that actually matters.

What changed

In April, Meta shipped Muse Spark.

It's their first closed model.

Not closed as in restrictively licensed. Closed as in there are no weights, hosted only, metered per token.

This is the company that made open weights a mainstream position.

Llama is the reason half the founders I know believed they'd always have an option that couldn't be taken away.

In July they opened a paid API for it, at roughly a quarter of what the American labs charge.

So the cheapest credible option on the market is now closed and metered.

From the company that spent two years arguing the opposite.

Alibaba did something quieter and, I think, more important.

They didn't reverse. They split.

Mid-tier Qwen models stay Apache 2.0. Genuinely open, download them today.

The frontier tier went API-only.

Qwen3.7 Max and the 3.8-Max preview have no downloadable weights.

Open weights are promised. They have not arrived.

I have a personal stake in that one.

We trained on Qwen.

That family is part of why I believed this bet was safe.

Read those two moves together and the pattern isn't 'open weights are dying.'

It's narrower and worse.

The frontier is going closed while the tier below it stays open.

You'll always be able to download something good.

The question is whether you'll be able to download the best thing.

The answer has been drifting toward no for four months.

Who's actually holding the line

Not the Western labs.

This is the part people get wrong when they treat it as an ideology fight.

On July 17, Moonshot released Kimi K3 at 2.8 trillion parameters, and called it the largest open-source model in the world.

DeepSeek V4-Pro tops the open leaderboards on agentic coding and graduate-level reasoning, under an MIT license.

That's about as permissive as it gets.

The open frontier right now is Chinese labs.

Not partly. Basically entirely.

Our own Forte runs on GLM 5.2 today for exactly this reason.

It was the best open-weight model when we tested, and it wasn't close.

I want to be careful here, because this is where the conversation usually turns into geopolitics and stops being useful.

I'm not making a claim about which country deserves this.

I'm telling you where the weights are, because that's the fact that determines what a small company can build on.

And I'd note the obvious risk in it.

An open frontier held by a small number of labs in one jurisdiction is not a durable open frontier.

It's a concentrated one that happens to be open right now.

That's a different thing.

I'd rather say so than pretend our bet is safer than it is.

Why this is load-bearing for us and not just interesting

Most people can read this as industry news.

I can't.

We killed our own hosted models earlier this year.

We were paying about $5,000 a month per GPU to serve small models to a small number of users.

Models a $600 endpoint could have served just as well.

That wasn't a mission. It was ego dressed up as infrastructure, and shutting it down was the right call.

But it means our answer to 'what runs your product' is now a bet on somebody else's release schedule.

I've said publicly why I like that bet, and I still do.

Because the weights are open, providers compete to serve the same model.

American inference companies serve GLM cheaper than the lab that made it.

Nobody can lock up the best open model, reprice it overnight, or take it away from us.

That argument is completely true.

And it only holds while the best open model exists.

If the frontier keeps closing, Forte's promise quietly degrades from 'the best model available' to 'the best model still being given away.'

Those are not the same product.

I'd have to say so out loud rather than let the wording carry me.

What I actually think happens

I don't think open weights collapse.

I think they stratify, and the gap becomes a subscription.

Here's the shape I'd bet on.

The tier below the frontier stays open and gets very good.

Good enough that for most real work, most days, the difference stops mattering.

That's already true for us. Most of what our team ships daily runs on open weights and nobody notices.

The frontier tier goes closed almost everywhere, because that's where the pricing power is and every lab eventually notices.

And the thing you pay for stops being the model.

It becomes the last few points of capability on the hardest problems.

Which, notably, is exactly how we already work.

Our own tooling handles the easy-to-mid tasks, and the hardest problems still go to the expensive frontier agents.

I've been running the future I'm describing for a month without calling it that.

The part I'd tell another founder

Don't pick a side on this.

Pick a position you can hold if you're wrong.

If your product's economics require open weights at the frontier, you have a dependency with no contract behind it, and the trend is against you.

That's not a reason to abandon it.

It's a reason to know which quarter you'd find out.

For us, the honest version is this.

We can absorb the frontier going closed, because our cost advantage comes from serving open models at cost plus a thin margin, and the tier that stays open is more than good enough for most of the work.

What we couldn't absorb is the open tier itself stalling.

A year with no meaningful open release while the closed models kept moving.

I don't think that's coming. Kimi K3 and DeepSeek V4-Pro shipped this month.

But I've now watched two labs I'd have called safe change their minds in a single quarter.

So I'm writing the assumption down, publicly, with a date on it.

That way when I'm wrong, there's a record of exactly what I believed and when.

Which is worth more to you than another confident take, and worth more to me than pretending I saw it coming.