the models · eu-hosted

the models you build with.

MakeMode builds on a small set of open AI models — from the quickest to the most polished. Pick by what you need; every one runs on European infrastructure, so your work never leaves the EU whichever you choose. Here's how they differ, and how they measure up — including speed numbers we measure ourselves.

try the demo →how they compare

pick by what you need

one default, three choices.

GLM 5.3 builds your page unless you pick another in the model menu. Pick Mistral if you want weights trained in Europe, or Qwen for the longest pages and a much longer wait. No wrong answer.

Light

Mistral Small 3.2

A short first draft, trained in Europe — for simple pages and quick iteration.

Mistral · France

technical detail →

The technical specs for this model — size, context window, licence and how it's served in the EU — will appear here.

Capable

Mistral Medium 3.5

The European flagship — detailed, well-built pages. A longer wait for higher-quality output.

Mistral · France

technical detail →

The technical specs for this model — size, context window, licence and how it's served in the EU — will appear here.

Default

GLM 5.3

The model that builds your page unless you pick another — the most polished output, and currently the fastest here too.

open weights · EU-hosted

technical detail →

The technical specs for this model — size, context window, licence and how it's served in the EU — will appear here.

Largest

Qwen3.5 397B

The biggest model here — the longest pages, and by far the longest wait (over two minutes in our latest run).

open weights · EU-hosted

technical detail →

The technical specs for this model — size, context window, licence and how it's served in the EU — will appear here.

how to choose in opencode → — in the web app it's a picker; in the terminal it's one line of config. Same models, one key.

how they measure up

the benchmarks, honestly.

Benchmarks say what a model is good at — not where your data goes (that's identical for all of them). Here's a snapshot on coding; rankings move weekly, so we link the live boards.

coding capabilitySWE-bench Verified · reported July 2026

Model	Best for	SWE-bench Verified
Mistral Small 3.2light · short drafts	quick pages, simple tools	general model
Mistral Medium 3.5EU flagship	detailed, polished pages	78%
GLM 5.3open weights, EU-hosted	most polished output	top open-weights †
Qwen3.5 397Blargest · open weights	deeply detailed builds	top open-weights tier †

On WebDev Arena (building web front-ends), Claude leads overall; among the open, EU-hostable models, GLM and the Qwen flagship tier rank highest.
† GLM 5.3 leads open-weight models on LMArena; its makers report 62% on the harder SWE-bench Pro (a tougher test, not comparable to Verified). See the live boards: LMArena · Artificial Analysis · SWE-bench.

measured on our rails

how fast, measured by us.

Leaderboards test models in a lab. This is what they do here — on the same European servers that build your pages. We run this benchmark ourselves, publish the raw data, and re-measure when the serving changes.

a full page, start to finish — same prompt to every modelmeasured 2 november 2026 · scaleway, paris · median of 3 streamed runs

GLM 5.3starts writing in ~0.2s · 219 tok/s — wrote the longest page of the fast models

14sfull build

Mistral Small 3.2starts writing in ~0.2s · 95 tok/s

16sfull build

Mistral Medium 3.5starts writing in ~0.1s · 61 tok/s

30sfull build

Qwen3.5 397Bthinks ~39s first, then 47 tok/s

136sfull build

How we measure: the same page-build prompt to every model, streamed, three runs, medians, no token cap — every build finishes naturally. "Full build" is wall time from request to last token; tok/s uses the provider's reported completion tokens (not chunk counting); "starts in" includes any thinking time, because that's the wait you actually see. Models write different amounts for the same ask — GLM wrote a ~3,200-token page where Mistral Medium wrote ~1,800, and Qwen ~4,400 — so time and detail trade off. Speeds drift with provider load and serving changes — we measured the same model swing 2× between days — so we re-measure rather than assume, and route each model to whichever EU provider serves it fastest. Raw data: model-speed.json.

the footprint

greener by where it runs.

Running in Europe isn't only about your data — it's about the grid underneath. MakeMode's models run on European data centres chosen for lower-carbon energy. We're measuring the energy and CO₂ of a real build the same way we measure speed — openly, with the raw numbers published here.

energy & co₂ per buildmeasured on our rails · published openly

Per-model energy and CO₂ figures are being measured now and will appear here — like the speed benchmark above, with the raw data linked. See also the full environment story.

the cost

what a build costs.

Open models are far cheaper to run than closed frontier services — and we'd rather show you the price than hide it. Here's the cost per model, so you can pick by budget as well as by polish, and see exactly what you're paying for.

cost per modelprice per 1M tokens · eu-hosted

Per-model prices are being finalised and will appear here, alongside a typical cost per build. See also inference economics and unit economics.

the same for every model

whichever you pick, it stays in the EU.

Sovereignty isn't a property of the model — it's how MakeMode runs all of them: on European clouds under GDPR (Scaleway in France, with some models served from EU data centres in Finland), your words in, a page out, nothing leaving the EU. An open model is a file we run, not a service phoning home — so where its makers are doesn't change where your data goes. That's true for Mistral and GLM alike.

the full sovereignty story →

build with one now.

Describe a page, watch it build, publish a link — on European rails.

try the demo →open the app