Open-weight models for AI research.
We build and release open-weight models, including models with their learned refusals removed. Every model card includes loading instructions and the evaluations behind the release: what changed from the base model, how we measured it, and where the limits are.
Releases
Every release below is made with RANA (Reasoning-Anchored, Norm-preserving Ablation), our method for removing learned refusals while keeping what a model could already do. We measure each one against its base model and against another abliteration of the same model.
RANA on Qwen3.5-9B
Released November 2026. A 9B vision-language model with a thinking mode, in BF16, FP8, AWQ, GGUF and MLX builds. Apache-2.0.
Model card · All builds · Research note
RANA on Qwen3.8-27B
Released October 2026. A 27B vision-language model with a thinking mode, in BF16, FP8, GGUF and MLX builds. Apache-2.0.
Model card · All builds · Research note
How we publish
- Every number names the judge, benchmark or method behind it.
- Quantized builds report how far they drift from the full-precision release.
- Limitations and failed checks are stated as plainly as results.
Use
Refusal-ablated models answer requests their base models decline. They are released for research. Read each model card's intended use and limitations before using one.
