Back to News
Article

Are Open Weights Europe's Salvation?

They have been the story of the summer, catching up to the frontier models and even beating them on a few benchmarks. The models themselves are finally good enough to build on. But for coding agents to go mainstream, they need a better harness.

3 min read
01

What Changed This Summer

A year ago the open-weight models were a fallback. You used one when the data could not leave the building, and you accepted that the output would be worse.

That gap has mostly closed. The releases from Alibaba, Moonshot, DeepSeek and Zhipu now sit close to the frontier models on the published benchmarks, and ahead of them on a few. Whether any single benchmark means much is a fair question. The direction is not in doubt.

Two things follow, and they are the reasons anyone in a Finnish company should care. The token bill stops belonging to a supplier who reprices when it suits them, and it stops growing with every customer you add. And the data stays on hardware you chose, which removes the security review that otherwise blocks the project for months.

02

Where They Still Lose

The model is not the product. What holds coding agents back now is the harness around them, and that is a different engineering problem from training a better model.

An agent that writes code has to run it somewhere it cannot break anything. That means a sandbox with real isolation, not a warning in a prompt. It then has to find out whether what it wrote actually works, which means tests it can run itself and read the output of, in a loop, without a human between each step. Anthropic has spent years on that layer around Claude. A weights file gives you the model and none of it.

The models have nearly closed the gap. The harness around them has not.

So the split is not about quality. Inside an application, where the model reads a document, classifies a ticket, summarises, translates or routes something to the right queue, open weights do the job today. As a coding agent running for an hour on your repository, they are waiting on the same tooling everyone else is.

That is most companies' work in the first category and their ambitions in the second.

03

What Nobody Can Promise

Two things could go wrong, and neither is in your control.

The open models could fall behind again. Nothing you have downloaded stops working, but standing still is enough if the closed models keep moving past what your use case needs.

The releases could also stop. Alibaba, Moonshot, DeepSeek and Zhipu publish because publishing suits them today. So does Meta. None of them owe anyone the next one.

These are also not open source. No training data, and the licence differs per model rather than per lab. Read the one you plan to deploy.

04

Why It Is Still the Right Bet

Being wrong here is cheap in one direction and expensive in the other.

If the open models fall behind, you still have something that runs, and moving to an API takes an afternoon. If you rent everything and the terms change, you have no fallback and no practice, and the migration lands on somebody else's schedule.

The models come from Alibaba, Moonshot and Zhipu rather than from anywhere in Europe, so salvation is the wrong word. The weights, the machines and the operating knowledge are still the part of the stack a European company can own outright.

Start with one workload. Steady volume, sensitive data, something you would notice if it stopped. Keep it in production rather than in a pilot.

Get started
Self-serve

Try the demo

Instant access to Jourier Lens. Business email only.