Eugeny 5.1 Vision: the merged dense policy, the honest visual boundary, and the v6 direction.
Eugeny — Sovereign European AI
A sovereign intelligence,
built for Europe.
Eugeny is a sovereign European AI, built and served under European jurisdiction by Fornace and aligned to the constitutional values of the European Union. Thirteen cognitive reflexes are trained into the weights rather than bolted on by prompts, and the goal is to compete with the frontier on real evidence — every capability claim on this site is tied to a receipt, or marked as planned, in development, or not yet verified.
The current product
Live model
5.1 Vision
Eugeny · BF16, unquantized
Serving context
262,144 tokens
Dense, single envelope
Vision gate
1 image / prompt
One video, sampled to two endpoint frames
Runtime
vLLM 0.25.0
Compiled CUDA graphs
Eugeny 5.1 Vision · merged dense policy · visual tower grafted from a pinned Qwen3.5-9B donor, non-destructive and reversible.
Four things to hold it to.
Eugeny is not a single feature. It is a set of commitments — the values it is aligned to and the goals it is built toward: sovereignty and EU constitutional alignment, instincts trained rather than instructed, context that ships now and reaches further in research, and a stated aim to compete with the frontier. Each pillar links to the page where it is substantiated.
Sovereignty & Compliance
Sovereign by design.
EU constitutional values are the alignment frame. No private personal data in the weights. Memory is external, sourced, scoped, and deletable.
Read more → Instincts in the WeightsReflexes, not rules.
Thirteen cognitive instincts at the data, policy, and code layers. Prompt rules hurt performance; weight-level instincts do not.
Read more → Long Context262,144 tokens today.
A dense serving envelope that ships now. HiLS-Attention is the learned-sparse research path toward millions of tokens.
Read more → Built to CompeteCompete with the frontier.
The v6 target is explicit: beat Sonnet 5 on PACE and the open-weight comparators on every comparable family. A goal, stated as a goal — not yet a verified result.
Read more →A sovereign alternative, built to compete.
Eugeny is a dense policy of roughly nine billion parameters, built and served under European jurisdiction by Fornace to compete on agentic and coding work while remaining open-weight. Sovereignty here is operational: where the model is served, what data enters the weights, and what it is allowed to claim are hard rules, not a label. No private personal data is baked into the weights, and the alignment frame is the constitutional values of the European Union — lawful political discourse and institutional criticism are answered, not refused, while exploitation, dehumanization, targeted abuse, minors, privacy invasion, and violence are met strictly.
The goal is a sovereign European alternative to the US frontier labs: open-weight, honest about its lineage, values-aligned, and built in Europe. The v6 program's explicit target is to beat Sonnet 5 on PACE and the open-weight comparators on every comparable family. That is a goal, stated as a goal — not yet a verified result.
The reflexes that govern how it works are trained, not instructed. An ETH Zurich 2026 study put the cost of prompt scaffolding at roughly twenty percent more tokens and three extra steps per task; Eugeny's answer is to bake thirteen cognitive instincts into the weights as trainable reflexes, so the right behaviour is the default rather than a fragile instruction.
Honesty is one of the values, not a marketing posture. Every capability claim on this site is tied to a receipt or marked as planned, in development, or not yet verified. The low multimodal score is published on the same surface as the wins, and the part of the predictor we do not trust is named first — because hiding any of it would breach the value, and because the goal is to compete on real evidence.
Where it stands today.
An honest accounting
The live model is Eugeny 5.1 Vision. It accepts at most one image per prompt, and a single video sampled to two endpoint frames. Multi-image reasoning is not verified: a twelve-image request transported, but its semantic gate scored nine of twelve, so the serving limit remains one. No image generation is claimed — that work belongs to a separate project, reported on its own.
The v6 research program — long context toward millions of tokens, and the process-awareness and mechanism-discovery instinct tracks — is in progress, not a shipped capability. Dense 256K context is what is proven and served today; everything beyond it is planned or open research, and is labelled as such on the pages that follow.
Where to go next.
Six pages. Each carries its receipts, or marks plainly what is not yet verified. Read in any order; none assumes the last.
Read the model card
How the instincts work
Thirteen reflexes trained into the weights, and the documented cost of prompt rules.
Sovereignty, memory, and tool honesty
European jurisdiction, external deletable memory, and no claim made without a trace.
The numbers and what they mean
PACE projections, direct visual proxies, and the predictor we do not trust.
The v6 research program
A verifier-first capability foundry. Dense 256K ships now; HiLS-Attention is research.
Who Fornace is
A sovereign European alternative, and the boundaries we hold to.