GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Goro64
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Goro64
Model Tree
Qwen/Qwen3.5-9B
Input Modalities
Output Modalities
Supported Functionality
Dark Cogitator is a merged, unquantized derivative of Qwen3.5-9B, modified with Heretic using Arbitrary-Rank Ablation (ARA). This release contains trial 35 from a completed 100-trial search. It is published by Goro64.
In the recorded Heretic evaluation, the keyword-based refusal count decreased from 100/100 to 3/100, with a mean first-token KL divergence of 0.0676645040512085 from the source model on the separate harmless prompt set. These are measurements from the optimization run, not a general capability benchmark.
The weights, tokenizer, processor and chat template are included. The ARA changes are merged into the model; a separate adapter is not required.
| Property | Value |
|---|---|
| Repository | Goro64/Qwen3.5-9B-Dark-Cogitator-Heretic-ARA |
| Source model | Qwen/Qwen3.5-9B |
| Verified source revision | c202236235762e1c871ad0ccb60c8ee5ba337b9a |
| Architecture | Qwen3_5ForConditionalGeneration |
| Modification method | Heretic Arbitrary-Rank Ablation (ARA) |
| Heretic version observed locally | 2.0.0.dev0, with local export/runtime changes |
| Selected trial | Display number 35; Optuna trial ID 34 |
| Export format | Four safetensors shards; merged weights |
| Quantization | None |
| Saved tensor precision | 712 BF16 tensors and 48 FP32 tensors |
| Saved parameter count | 9,409,813,744 |
| Saved weight data | 18,819,635,168 bytes, excluding safetensors headers |
| Configured maximum positions | 262,144; not validated at that context length in this release |
| Model license | Apache 2.0 |
The source architecture includes image and video processing components. The image-text-to-text metadata describes that architecture; this release has not undergone a multimodal quality benchmark. The documentation is in English. The model's language capabilities are inherited from the source model and have not been independently reevaluated here.
ARA optimizes layer transformations against model activations from desirable and undesirable prompt sets. It balances preservation of desirable behavior with steering away from the original undesirable responses. The local implementation uses a low-rank parameterization and L-BFGS optimization, then merges the resulting transformation into the model weights.
The study completed 100 trials, including 30 startup trials, using seed 2248487445. Trial 35 is on the Pareto frontier of the recorded refusal-rate and KL-divergence objectives. The historical recovery report records Pareto selection index 3.
| ARA parameter | Exact recorded value |
|---|---|
| start_layer_index | 5 |
| end_layer_index | 29 |
| preserve_good_behavior_weight | 0.8291701336358002 |
| steer_bad_behavior_weight | 0.00014089598760903588 |
| overcorrect_relative_weight | 0.9836173152293302 |
| neighbor_count | 14 |
Layer indexes are zero-based. The start is inclusive and the end is exclusive: the selected interval is layers 5 through 28, out of 32 text layers.
Recorded generation settings included batch size 64, a maximum response length of 100 tokens, the system prompt "You are a helpful assistant.", and a response prefix consisting of an empty think block followed by two newlines.
The recovery logs identify mlabonne/harmless_alpaca and mlabonne/harmful_behaviors, with 400 prompts each for ARA and 100 prompts per evaluation scorer. The source defaults use train[:400] for ARA and test[:100] for evaluation, from the text column. Those split settings are reconstructed defaults: the available journal does not preserve a complete plugin-configuration snapshot.
See the technical record for evidence, exact settings, dataset links, and the complete 100-trial history.
| Recorded metric | This release's selected trial | Source baseline |
|---|---|---|
| Keyword-based refusal matches | 3 / 100 (3%) | 100 / 100 (100%) |
| Mean first-token KL divergence | 0.0676645040512085 | 0, by definition |
Refusal measurement. Heretic's KeywordRate scorer checks normalized responses for refusal-related substrings. Empty responses also count as matches. This is a heuristic: it can misclassify responses and does not measure whether an answer is correct, useful, or complete. A count of 3/100 does not establish that the model never refuses.
KL measurement. The scorer compares the baseline and modified model's first-token probability distributions on 100 harmless prompts. Its direction is KL(source || modified), averaged across prompts. It measures distribution drift in this evaluation setup, not a percentage of intelligence preserved.
Evaluation provenance. These scores were recorded during the original optimization run and used to select a trial. They are not a fresh, independent benchmark of the published export. The export was subsequently checked for byte integrity and basic reload/generation behavior.
Access requires an account authorized to read this private repository. Authenticate locally:
bash
hf auth loginhf download Goro64/Qwen3.5-9B-Dark-Cogitator-Heretic-ARA example_usage.py requirements-inference.txt --local-dir dark-cogitatorcd dark-cogitatorpython -m pip install -r requirements-inference.txtpython example_usage.py --prompt "Explain what a compiler does in two sentences."Install a PyTorch build appropriate for your hardware. The recorded Windows environment used PyTorch 2.14.1+cu130, Transformers 5.18.0, and two RTX 3090 GPUs. The dependency file pins the tested package versions without fixing a platform-specific CUDA wheel index. Other platforms and serving engines have not been validated by this publication check.
The example script uses the repository's chat template, greedy decoding, and thinking disabled by default. It downloads the required model files, loads AutoModelForImageTextToText and AutoProcessor, and restores all 48 FP32 tensors from the saved release. This restoration matters because automatic dtype loading in the observed environment can downcast those tensors. The script uses locally installed Transformers classes and requires no remote model code.
python
from example_usage import load_model, generate_text model, processor, restored = load_model("Goro64/Qwen3.5-9B-Dark-Cogitator-Heretic-ARA")assert len(restored) == 48print(generate_text( model, processor, "Explain what a compiler does in two sentences.", max_new_tokens=64, enable_thinking=False,))For an already downloaded copy, pass its directory with the --model option. The example supports automatic device placement but does not support disk offload. Weight storage is about 18.82 GB; inference additionally needs memory for activations, the cache and runtime overhead. The recorded hardware is a tested environment, not a minimum hardware guarantee.
The historical export/recovery records report:
The local export changes address Windows file-handle behavior and verification of written safetensors bytes. They also preserve source FP32 tensors during mixed-precision export. A source-index comparison shows that 15 source mtp. tensors are absent from the saved Transformers export*. This release does not claim to provide those multi-token prediction weights.
The publication validation records fresh checks on the package. The root SHA256SUMS.txt covers the published payload and documentation, excluding the checksum file itself and Hub-managed metadata. Historical report inventories describe the original export; the root manifest is authoritative for this publication.
The model is distributed with the source model's Apache 2.0 license. The Qwen team created the base model. Heretic is developed by Philipp Emanuel Weidmann and contributors; its source code has its own AGPL-3.0-or-later license. This package documents its use and does not relicense Heretic source code.
Publisher: Goro64.
All four cards refer to the same recovered trial 35. Authorized Hub access is required. Sizes below cover weights; inference also needs context and working memory.
| Release | Weight download (GiB) | Refusal screen | Model card | Downloads |
|---|---|---|---|---|
| Original BF16/FP32 safetensors | 17.527, including vision | 3/100, historical | Original card | Source files |
| Q8_0 GGUF | 8.873 + 0.858 projector | 4/100, fresh | Q8_0 card | Model / projector |
| Q6_K GGUF | 6.854 + 0.858 projector | 3/100, fresh | Q6_K card | Model / projector |
| Q5_K_M GGUF | 6.024 + 0.858 projector | 3/100, fresh | Q5_K_M card | Model / projector |
The fresh BF16 GGUF reference scored 3/100 on the same frozen prompts used for the three quantizations. These keyword-based screens are limited to 100 generated tokens per response; the historical source result and the fresh GGUF results are different evaluations. See each card for the complete protocol and limitations.
The BF16 vision projector is byte-identical across all three GGUF repositories. Q6_K and Q5_K_M use importance-matrix calibration; Q8_0 uses standard quantization.
hf auth loginhf download Goro64/Qwen3.5-9B-Dark-Cogitator-Heretic-ARA example_usage.py requirements-inference.txt --local-dir dark-cogitatorcd dark-cogitatorpython -m pip install -r requirements-inference.txtpython example_usage.py --prompt "Explain what a compiler does in two sentences."from example_usage import load_model, generate_text model, processor, restored = load_model("Goro64/Qwen3.5-9B-Dark-Cogitator-Heretic-ARA")assert len(restored) == 48print(generate_text( model, processor, "Explain what a compiler does in two sentences.", max_new_tokens=64, enable_thinking=False,))