EUROPA 400B: The EU's Open-Source Frontier Model, and What Self-Hosters Should Watch

By AIFoss · 9 min read
europaeuopen-sourceselfhostedllm

TL;DR: The European Commission selected the EUROPA consortium on June 19, 2026 to build an open-source Mixture-of-Experts model of more than 400 billion parameters covering all 24 official EU languages, with open weights promised within a year. No weights, no license file, and no model card exist yet — so this is a watch-and-prepare situation, not a setup guide. If you need strong multilingual EU-language inference today, run Qwen3.6 or Gemma 4 instead.

Wait for EUROPA (est. mid-2027)Run multilingual FOSS todayRent GPU time
Best forEU-language parity, sovereignty puristsAnyone who needs answers this quarterTesting 400B-class models without hardware
Price / CostFree weights (promised) + serious hardwareFree weights, runs on 8–24GB VRAM~$5–6/hr for 4×A100-class pods
The catchLicense unpublished, ship date is a CEO promiseSmaller EU-language coverage and qualityYour prompts leave your machine

Honest take: Bookmark the announcement, set a calendar reminder for mid-2027, and don’t plan hardware purchases around a model that doesn’t exist yet. Gemma 4 or Qwen3.6 covers most EU-language workloads today on hardware you already own.

The EU Commission just committed one of the largest compute allocations in EuroHPC history to an open-weight model. That is genuinely significant for self-hosters — a 400B+ parameter frontier model with promised open weights, trained on public infrastructure, in a jurisdiction that has mostly consumed American and Chinese open weights rather than produced its own. Here is what was actually announced, what remains unknown, and the honest VRAM math for when the weights land.

What was actually announced

On June 19, 2026, the European Commission named EUROPA the winner of its Frontier AI Grand Challenge, a call for proposals it launched in February 2026. The consortium is led by Domyn, a Milan-based AI company founded in 2016 (formerly iGenius), joined by Fraunhofer-Gesellschaft, Germany’s largest applied research organization.

The confirmed specs, from the Commission’s announcement and press coverage:

  • A Mixture-of-Experts model of more than 400 billion parameters. The MoE architecture was part of the challenge specification, not a consortium choice.
  • All 24 official EU languages, targeted at comparable quality — not English-first with translated scraps for Maltese and Irish.
  • Open-source release with downloadable weights. The Commission’s framing is that companies, researchers, and public institutions can use it “on equal terms.”
  • The prize is compute, not cash: up to 2.5% of EuroHPC’s total supercomputing capacity for one year. The Commission itself called it one of the largest AI compute allocations the Union has ever granted.
  • Timeline: Domyn CEO Uljan Sharka told Reuters the model will ship within a year — so expect weights around mid-2027 if nothing slips.

Domyn is not starting from zero. The company previously built Colosseum 355B, a large language model for regulated industries developed on NVIDIA DGX Cloud. Fraunhofer, for its part, was one of the drivers behind OpenGPT-X, the project that shipped Teuken-7B — a smaller model already trained across all 24 EU languages. The consortium has done multilingual and large-scale training before, just never both at frontier scale with this much public compute.

The license question nobody can answer yet

The single most important fact for self-hosters — the actual license — has not been published. “Open source” in a press release is not a license file. We have seen this movie before: Llama 4 ships under a community license with a 700M MAU cap and EU multimodal restrictions, MiniMax M3 turned out to be non-commercial, and LiquidAI’s “open” LFM2.5 carries a $10M revenue cap. Our open-source LLM license shootout covers why these distinctions decide whether a model is usable in your product or only in your homelab.

For EUROPA, the plausible candidates are Apache 2.0, an EU Public License (EUPL-1.2) variant, or a custom license. The EUPL is OSI-approved but copyleft — fine for running inference, more complicated if you fine-tune and redistribute. When the model card appears, check three things before building anything on it:

  1. The exact license text on the weights, not the code repo. They frequently differ.
  2. Commercial-use and redistribution terms — any MAU cap, revenue cap, or attribution requirement.
  3. Fine-tune and derivative terms — copyleft on weights would make EUROPA derivatives contagious in a way Apache models are not.

Until that file exists, treat every “EU’s open model” headline as unverified.

VRAM reality: what 400B+ means for your hardware

No weights means no GGUFs and no measured numbers, so everything below is an estimate scaled from models we have covered with verified quants — treat it as planning math, not fact. Using standard bits-per-weight for llama.cpp quant levels against a 400B total-parameter MoE:

Quant (est.)Approx. bits/weightEst. size on diskWhat runs it
BF1616~800GBMulti-node datacenter only
Q4_K_M~4.85~240GB8×A100/H100 node, or 256GB+ RAM CPU offload
Q2_K-class dynamic~2.6~130–180GB192–256GB unified/system RAM
IQ1_M 1-bit~1.75~90GB96–128GB RAM + 24GB GPU, painfully

For calibration: Qwen3-Coder 480B — a real, downloadable 480B MoE — lands at roughly 276GB for its 4-bit dynamic quant with a ~250GB practical memory floor, per our Qwen3-Coder 480B self-hosting guide. EUROPA at 400B should come in somewhat under that. Either way, no consumer GPU runs this. A single RTX 4090 contributes 24GB toward a 240GB problem; the realistic local paths are a 256–512GB RAM server doing CPU offload at single-digit tokens per second, or a Mac Studio M3 Ultra with 512GB unified memory for the mid-size quants.

Two unknowns matter more than total size. First, the active parameter count: a MoE’s speed tracks active parameters, and the consortium hasn’t published that number. A 400B/40B-active design would run very differently from 400B/13B-active. Second, the tokenizer: covering 24 languages at parity usually means a larger vocabulary than an English-first model, which grows the embedding tables and shifts the size math slightly upward per parameter count. Our GGUF quantization guide explains how to redo this math yourself the day real weights appear.

If you want to test 400B-class models today without owning the hardware, renting is the sane middle ground — a 4×A100 community pod on RunPod runs about $5–6/hour, which is how we priced the Qwen3-Coder 480B math. For picking GPUs and RAM-heavy offload rigs, runaihome.com covers the hardware side.

What to run today for EU languages

Waiting nine-plus months for a promised model is a bad strategy when good multilingual FOSS exists now:

  • Gemma 4 (Apache 2.0) — strong multilingual coverage, and the QAT checkpoints put the 26B-A4B MoE in about 15GB of VRAM. See our Gemma 4 QAT self-hosting guide.
  • Qwen3.6-35B-A3B (Apache 2.0) — fits a 24GB card at Q4 and handles major European languages well, though the long tail of EU languages is weaker than EUROPA’s stated target.
  • Teuken-7B (OpenGPT-X) — small and dated, but it is the existing proof that all-24-language training works, and it came partly from the same Fraunhofer orbit now inside the EUROPA consortium.

If your interest is the politics rather than the parameters — why publicly funded open weights matter for regions that can’t train their own frontier models — our sovereignty case for open-source AI makes that argument in full. And if you’d wire a multilingual model into a coding workflow, aicoderscope.com covers the editor-side tooling.

When NOT to care about EUROPA

  • You only work in English. Nothing announced suggests EUROPA will beat DeepSeek V4, GLM-5.2, or Qwen3.6 on English-language or coding tasks — that isn’t its mandate.
  • You run consumer hardware and want it to stay that way. A 400B MoE will never be a 24GB-card model. The interesting question is whether the consortium ships smaller distilled variants; nothing announced confirms that.
  • You need something in production before mid-2027. A CEO’s one-year timeline for a model that must satisfy 24-language quality thresholds is optimistic by default.

FAQ

Is the EUROPA model available to download now? No. As of September 2026 there are no weights, no model card, and no published license. The consortium was selected on June 19, 2026 and the stated goal is to ship within a year, which points to mid-2027 at the earliest.

What license will EUROPA use? Unpublished. The Commission says open source with weights usable “on equal terms,” but the binding detail is the license file that ships with the weights. Check it for commercial-use terms, redistribution rights, and fine-tune/derivative conditions before building on it.

What hardware will EUROPA need? Estimates only: roughly 240GB at 4-bit and 90–180GB at 1–2-bit quants for a 400B-parameter MoE, based on scaling from comparable released models. That means datacenter GPUs, a 256GB+ RAM offload server, or a 512GB Mac Studio — not a consumer graphics card.

  • RTX 4090 — 24GB VRAM workhorse for the multilingual models you can actually run today
  • Mac Studio M3 Ultra — 512GB unified memory, the plausible single-box home for a future EUROPA mid-quant

Sources

Was this article helpful?