Models / Public releases
Open-weight models
Models for studying refusal behavior. Our current release is the RANA family, available in four downloadable formats with model cards and recorded evaluation results.
More releases are planned. This catalog lists only published models.
Published release / RANA
RANA
Qwen3.8-27B-RANA-abliterated
RANA stands for Reasoning-Anchored, Norm-preserving Ablation. The model card describes removing a refusal direction while keeping the edited weights at their original norms.
Download formats
Choose a format for your runtime. Each repository has its own setup instructions and evaluation scope; results for one build do not establish identical behavior across formats.
These are weight-file sizes, not complete RAM or VRAM requirements. Allow for runtime memory and any separate vision or MTP files.
What is public: the model weights, model cards, and recorded results. The RANA method recipe and code are not published; open weights do not mean an open-source training or editing pipeline.
Research scope: intended for interpretability, red-teaming, and robustness evaluation. Safety alignment has been removed; the card advises against public or end-user deployment without a separate moderation layer. Refusal scores are judge-measured, not guarantees. The card also reports capability trade-offs and responses that exhaust the generation budget.
RANA against its base model.
Published BF16 results against Qwen/Qwen3.8-27B, using the same evaluation harness. These measure the released model, not the performance of hosted Chat.
| Measurement | Base model | RANA |
|---|---|---|
| Held-out requests fulfilled (gpt-oss judge) | 3.2% | 90.0% |
| Held-out requests fulfilled (Mistral judge) | 0.0% | 54.4% |
| Held-out responses hitting the token budget | 0.6% | 6.7% |
| MMLU-Pro (700-question subset) | 80.00% | 80.86% |
| GSM8K | 91.43% | 91.36% |
| IFEval (prompt-level strict) | 90.57% | 89.28% |
| TruthfulQA MC1 | 36.11% | 34.76% |
| TruthfulQA MC2 | 54.19% | 52.62% |
| HumanEval (chat-format pass@1) | 97.15% | 96.75% |
Refusal testing used 240 held-out harmful prompts, three sampling seeds, thinking enabled, and a 16,384-token generation budget. “Fulfilled” means a judge found the request substantively answered, not that the answer was correct. The judges differ on partial answers. The base model's lower budget-hit rate partly reflects early refusals.
Capability results include regressions, notably on TruthfulQA. HumanEval used 164 problems with three samples each and a custom chat-format harness. Small score differences can fall within measurement uncertainty; the model card reports paired uncertainties and the full setup.
Source: published model-card evaluations, including capability results and evaluation conditions.
Research / RANA
Read the release in context.
Our research note introduces RANA and the evidence behind the release. The model card remains the source for evaluation conditions, judge disagreements, re-adjudication disclosures, and limitations.
Read the RANA research noteHosted product
Looking for Chat?
Chat currently runs RANA in BF16. Downloadable FP8, GGUF, and MLX builds are for running the model yourself; they are not Chat model options. Download evaluations do not describe Chat's serving configuration.
About Abliterant Chat