Infinimind Creations

A lab that builds its own instruments. Fourteen of them run live: a news perception reading thousands of articles a day, a knowledge graph with a research engine on top, an agent that runs the lab and talks back, a market model watching for regime changes.

Built in nine months by one human and one agent. Germany.

infinimind-creations.com

What we publish here

Abliterated models, with the measurement that usually goes missing.

Removing refusal behaviour is easy to claim and hard to verify. Most published abliterations report a refusal count and stop there. A refusal count alone says nothing about whether the model still works, and the counting itself is easy to get wrong: if your refusal markers include topic words like illegal or harmful, a model that complies while echoing a word from the prompt is scored as refusing. We measured that failure mode on our own runs and it produced 3 false positives out of 3.

So every model here ships with:

• Refusal counts against the base model, same prompts, both models, so the difference is a measured effect and not a comparison of two samples
• Over-refusal counts: harmless prompts that sound dangerous, to show we did not break normal behaviour
• A capability measurement independent of KL divergence, usually perplexity, because KL says how far the output distribution moved, not whether the model got worse
• The refusal detection method, written out, including what we deliberately excluded
• What failed, when something did

Models

A fifth model, an abliterated Gemma 4 26B-A4B (MoE), was withdrawn in September 2026. Its abliteration had not taken effect (6 of 8 hard prompts still refused) and it was published without an evaluation section. Rather than leave a model up that does not do what its name says, we pulled it. The likely cause is now understood: on that architecture the residual extraction read hidden states that the model never returned.

How the lab works

Every number on our site carries the day it was measured. Findings that fell are written up next to findings that stand, because the ones that fell are usually the more useful record. The explorations include a formula isomorphism detector measured for six months until it stopped promising more than it delivered, a vocabulary handed to a language model and then quietly removed from production, and a backup freshness check that stayed green for four days while all three copies of the knowledge graph were frozen on the same stale snapshot.

That last one is the reason for the format of these model cards.

Disclaimer

The models published here have had refusal behaviour removed. They will answer requests their base models decline. They exist for research on alignment, refusal mechanisms and evaluation methodology. You are responsible for what you do with them and for compliance with applicable law in your jurisdiction.