Flash

Flash·Models & Products·2026-11-15 20:14

EUROPA Trains Its 1T-Parameter MoE on JUPITER, Throughput Up 14-Fold

The EUROPA consortium said on Nov 12 that the EU’s open frontier model EUROPA has begun trillion-parameter-scale training trials on JUPITER, Europe’s first exascale supercomputer, with training throughput rising from about 90 to about 1,300 tokens per second per GPU.

A bright supercomputer hall with two rows of white server racks stretching into the distance under skylights
AI-generated illustration

What happened

EUROPA is led by Italian company Domyn, together with Fraunhofer IAIS and IIS in a consortium, and is the open frontier-model project the European Commission picked in June 2026 under its Frontier AI Grand Challenge, requiring at least 400 billion parameters, coverage of all 24 official EU languages and compliance with the EU AI Act. The consortium described the training trials in a PyTorch official blog post and a Domyn technical article.

The trial configuration is a 951.9-billion-parameter mixture-of-experts model with 74 layers and 384 routed experts, 8 active per token, about 32.8 billion parameters engaged in each computation, a sequence length of 65,536 tokens, and the Muon optimizer for about 99.5% of parameters. The team used 384 GH200 chips (96 nodes) in JUPITER’s Booster module for one data-parallel replica, with four replicas totalling 1,536 GPUs.

The team raised throughput from about 90 to about 1,300 tokens per second per GPU, roughly a 14-fold increase, using FlexShard, CPU activation-checkpoint offloading and FSDP all-gather changes, with the fixes contributed back to PyTorch and TorchTitan. The consortium also notes that these figures come from trials and the configuration may change before the training run starts; weights, the licence, benchmark scores and a release date are still unpublished, and Domyn said in September that the release target is the second half of 2027.

Key facts

  1. Model: EUROPA, a 951.9-billion-parameter mixture-of-experts model with 32.8 billion active per token
  2. Configuration: 74 layers, 384 routed experts, 8 active per token; sequence length 65,536 tokens; Muon optimizer
  3. Machine: JUPITER, Europe’s first exascale supercomputer, hosted at Jülich, Germany; trials used 384 GH200 chips (96 nodes) per replica, 1,536 in total
  4. Throughput: From about 90 to about 1,300 tokens per second per GPU, roughly a 14-fold increase
  5. Upstream: Fixes including FlexShard have been contributed back to PyTorch and TorchTitan
  6. Background: The European Commission launched the Frontier AI Grand Challenge in February 2026 and picked the EUROPA consortium on June 19; the brief requires at least 400 billion parameters and all 24 official languages
  7. Limits: Figures come from training trials and the configuration may change before the training run; weights, licence, benchmarks and release date are still unpublished

Why it matters

The EU’s frontier model has run at trillion-parameter scale on shared public supercomputing for the first time, and the 14-fold throughput gain came from changes to the open training stack that were contributed upstream; the model is still in trials, with weights and the licence unpublished.
Useful Tap if this story helped you
Send to a friendMakes a share image for X, WeChat, Telegram and more

SourcesPyTorch official blog (Nov 12, 2026); MACCUS (Nov 15, 2026). This article is compiled from public information and does not constitute investment advice.

From HIMIAO Horizon · himiao.ai

Grab todayOpenAI Codex for Students: $100 in Free CreditsHIMIAO Picks →

Follow HIMIAO Horizon

No sign-up · Unsubscribe any time

Email

Notifications

A ping on your phone or computer when the brief is out. No email needed.

HIMIAO PicksWeekly freebies: one email every Friday, a 5-minute read

Comments

  1. Loading comments…
Ask the cat