Giotto's free reasoning model is live — its license bans benchmarking
Switzerland's Giotto enters the frontier race with a free, EU-hosted flagship and a one-GPU model — as an open-weight VLM booster lands with no such strings.
Copy markdown
A free flagship reasoning model, hosted in Europe
Swiss lab Giotto — EPFL math alumni who placed 2nd in the 2025 ARC-AGI competition — made its flagship model, Alto, free for everyone: anonymous web chat at giotto.ai plus a free shared-cloud API (capped at one concurrent request). It runs on eight B300s, is GDPR-aligned, and routes around US cloud entirely.
The catch: you're barred from benchmarking it
Alto's terms forbid reverse-engineering its weights or architecture and bar publishing any benchmark of the model without Giotto's written consent. You can call it today, but you can't independently verify its reasoning claims — factor that into any eval or procurement decision.
Solo: 'the smartest model to run on one GPU'
Giotto's sibling model, Solo, is built to run on a single GPU — a workstation on-prem or a dedicated cloud instance in Switzerland or the EU, licensed per GPU. For teams with data-residency rules, it's a self-hostable reasoning option that never touches a US provider.
ZwZ2: open-weight VLMs that 'connect evidence'
inclusionAI released ZwZ2 in 4B, 9B, and 27B sizes — weights, GGUF builds, and the RL-VQA dataset all on Hugging Face. Trained with its EvidenceConnect pipeline, each size beats its own base VLM on all 25 perception benchmarks, and the 4B fits a small GPU. Unlike Giotto, no license strings — download and run.