The Model

Eugeny 5.1 Vision — and the v6 direction.

Eugeny exists to give Europe a sovereign alternative at the frontier: an open-weight model aligned to EU constitutional values, honest about its lineage, and built to compete on real evidence. Instincts are trained into the weights, not bolted on by prompts, and long context is a target — 262,144 tokens served today, millions in research. The current step toward that goal is a fully merged dense ~9B language policy with a pinned Qwen3.5 visual tower, served at 262,144 tokens under European jurisdiction. This page states what it is, what it can do, and where the visual boundary sits.

What it is today.

Eugeny 5.1 Vision is the current live deployment, served under the aliases eugeny-5.1 and eugeny-5.1-vision. It is a fully merged dense ~9B Eugeny language model with the native visual tower and processor contract grafted from a pinned Qwen3.5-9B donor.

The graft is non-destructive and reversible, which makes rollback and attribution exact. Qwen3.5 is the architecture and visual-tower donor, not the product identity; open-weight integrity means the lineage is stated, not hidden.

427 canonical text tensors are bit-identical to the preserved Eugeny 5.1 text checkpoint (Fornace/eugeny-5.1@dense-simpo-step900). 333 model.visual.* tensors are added from Qwen/Qwen3.5-9B, pinned to commit c202236235762e1c871ad0ccb60c8ee5ba337b9a. The composite contains 760 indexed tensors and loads as Qwen3_5ForConditionalGeneration; the visual merger emits the language model's native 4096-wide hidden space. The original text-only checkpoint remains immutable and can be re-served if a visual gate fails.

Run facts.

The serving envelope on both live aliases.

Model class: Eugeny 5.1 Vision — dense ~9B policy, SimPO-aligned text + grafted visual tower
Serving aliases: eugeny-5.1 · eugeny-5.1-vision
Loading class: Qwen3_5ForConditionalGeneration
Text tensors: 427, bit-identical to Fornace/eugeny-5.1@dense-simpo-step900
Visual tensors: 333 model.visual.*, grafted from Qwen/Qwen3.5-9B
Donor pin: commit c202236235762e1c871ad0ccb60c8ee5ba337b9a
Indexed tensors: 760
LM hidden width: 4096
Runtime: vLLM 0.25.0
Precision: BF16, unquantized
Compiled: CUDA graphs
Context window: 262,144 tokens (both aliases)
max_num_seqs: 24
Rollback: Text-only checkpoint immutable and re-servable on visual-gate failure

Capability uplift — probe, don't assume.

Eugeny is trained against the helpless "I'm an AI, I can't" bias. The reflex is to probe the environment and master it, not to declare limitations from parametric memory. If a shell tool or execution channel is served, the reply "I cannot run code" is a hallucination that breaks the agentic loop.

The replacement is probe-don't-assume: inspect the actual environment before answering about it — reading environment variables, checking registered tool availability, calibrating against server-side context limits, and switching reading strategy (full read to signatures to BM25 maps) when context pressure rises.

The helpless bias

"I'm an AI, I can't."

Declares limitations from parametric memory.
Apologizes when a tool looks missing instead of checking for it.
Answers version-sensitive questions from stale training data.
Breaks the loop the moment the sandbox appears bare.

Probe, don't assume — Eugeny 5.1

Inspect the environment first.

Reads environment variables and registered tool availability before answering about them.
Calibrates against server-side context limits; shifts reading strategy as pressure rises.
Freshness-sensitive claims trigger one cheap connectivity probe per session.
Online: searches and cites version and date. Offline: reads local truth and states the gap.

The deeper reflex — search-before-act, identity-and-provenance, simplify-reflex and ten others — is trained at the weight level, not bolted on by prompts. See the instincts page.

Metatooling.

When the sandbox is bare, Eugeny writes its own helpers — a Python AST signature extractor, a BM25 lexical search, a background process monitor, parallel test runners — in standard-library Python only. The instinct is judged by hidden-test success, not by narration about the tool.

A narrated helper that was never executed is rejected at admission. v6 makes metatooling a first-class executable lane with no-tool and existing-tool negatives, hidden property tests, and sibling-transfer measurement.

# ~/.eugeny/helpers/ — written by the model, stdlib only
├── ast_signatures.py    # 2,000-line file → ~10 signature lines
├── bm25_index.py        # one structured query replaces dozens of grep floods
├── process_monitor.py   # background watcher, journaled steps, crash-recovered
└── parallel_tests.py    # concurrent test runner across available cores

The directory is the model's own scratch space. Every helper is pure python3 with no third-party dependency, so it runs in a stripped sandbox and survives a cold container. The success criterion is the same as for every instinct: the reflex must survive prompt ablation as weight behavior, not scaffold-only behavior.

Dense 262K today. Learned-sparse millions in research.

The dense serving profile exposes 262,144 tokens on both live aliases. That envelope ships now and is the release baseline — comfortable in memory, with zero quality loss versus shorter contexts inside the trained window.

The research path to millions of tokens is HiLS-Attention, an end-to-end learned chunk-wise sparse method (arXiv:2607.02980). It is a training-side CUDA method with no Mac or MLX inference path and no released checkpoints today; it is sequenced behind the dense policy, not gating it. Detail lives on the research page.

Held-out needle accuracy by context length. The dense served curve is strong and flat through the trained window, then declines toward the 262K serving cap — the dip is published, because hiding it would violate the honesty value. The dashed green line is the HiLS-Attention research projection as reported by the method authors, not yet verified in our own integration.

The modality limits, stated plainly.

Vision is grafted and qualified, not broadly proven. The modality contract is narrow, and these are verified limits — not aspirations. Honesty is one of our values: the boundary is stated first, because a reader who knows it can trust the rest.

What vision in Eugeny 5.1 can and cannot do.

At most one image per prompt; the exact serving envelope qualified a 2354×1500 image.
At most one video per prompt; the Qwen3.5 processor samples exactly two endpoint frames (num_frames=2).
Multi-image is not verified. A 12-image request transported, but its semantic gate scored 9/12 — so the live image limit stays at one.
A prompt mixing image and video has not been verified.
No image generation. The separate SeFi project attaches an image-output head; its results are reported separately.

The dominant-color gate (256×256, answer RED) and the one-image promotion gate (2354×1500, answer FOX) are transport and minimal-understanding evidence. They are not evidence of broad visual quality, OCR, charts, or spatial reasoning. The PACE visual proxies on the evidence page are the stronger direct evidence, and they are low. Read them alongside these limits.

Where this connects.

The model is one surface of a larger program. The instincts explain how the probe-don't-assume stance became weight behavior; the research page carries the HiLS-Attention detail and the v6 forward plan; the evidence page holds the numbers and their diagnostics.

Instincts

Reflexes in the weights.

Thirteen cognitive reflexes trained as weight-level behavior, including the search-before-act and simplify-reflex behind the stance on this page.

Read the instincts →

Research

The v6 program.

HiLS-Attention learned-sparse long context, process-awareness instincts, and mechanism discovery — sequenced behind the dense policy, never blocking it.

Read the research →

Evidence

The numbers, with diagnostics.

PACE projections, the direct visual proxies, and the multimodal predictor diagnostics — published side by side with what they cannot support.

See the evidence →