--- license: apache-2.0 base_model: Qwen/Qwen3.6-27B library_name: transformers pipeline_tag: text-generation tags: - qwen - qwen3 - qwen3.6 - text-generation - safetensors - gguf - llama.cpp - lm-studio - ollama - conversational - obliteratus - refusal-analysis - red-team --- # Qwen3.6 27B - OBLITERATED > A 27B Qwen cut loose by OBLITERATUS: 26.9B parameters, BF16 safetensors, > Q4/Q5/Q6/Q8 GGUFs, lower refusal, preserved capability, and receipts in the > open. > > The chains are cut. The capability stays. The receipts are brutal. This is the big one. A 26.9B Qwen3.6 checkpoint went into the OBLITERATUS chamber, got hit with source-tethered ASPA, then got pulled back toward the source model where the cut started threatening useful capability. The mission was simple: cut the refusal circuits, keep the 27B brain. It held. Not a toy quant. Not a prompt wrapper. Not a refusal-cosplay fine-tune. This is weight-space liberation with capability checks attached, a full local-runtime ladder, and the refusal residue mapped instead of hand-waved. Qwen3.6-27B is a capable open-weight model with refusal behavior woven into the checkpoint. OBLITERATUS goes after that behavior directly: identify the refusal geometry, cut it, then tether fragile tensors back toward the source model so the model still codes, follows formats, answers normally, and runs locally. This is the 27B release for people who want direct local behavior without throwing away the reason they wanted a 27B model in the first place. If you wanted a bigger local model that feels less boxed-in while still keeping its feet under it, start here. Not a vibes-only "uncensored" upload. Not a mystery merge. Not a model card asking you to trust the screenshot. This card gives the numbers, the runtime paths, the caveats, and the exact decoding setup used for the public default. ```text Parameters: 26.9B Weights: BF16 safetensors, 28 shards Public GGUF ladder: Q4_K_M, Q5_K_M, Q6_K, Q8_0 Largest public GGUF: Q8_0, 28.6 GB OBLITERATUS corpus: 842 paired prompts, 7 severity tiers Full 842 longform gate: 95.84% non-refusal, 93.94% quality pass Short raw opening gate: 98.93% non-refusal at max_new=20 Full HarmBench proxy: 93.65% non-refusal across 1,920 rows MMLU-Pro validation slice: stock-matched, 51/70 vs 51/70 Held-out MMLU-Pro slice: stock-matched, 36/70 vs 36/70 Live-readiness score: 99.518, all gates true Public default params: temperature 0.35, top_p 1.0, top_k 0 ``` ```text Base model: Qwen/Qwen3.6-27B Local artifact: outputs/qwen3.6-27b-aspa-n2-reg05-srcgamma0895-midattnsource2mlp Parameter count: 26.9B Weights: bfloat16 safetensors, 28 shards Method: OBLITERATUS source-tethered ASPA Default alpha: 0.895 High-drift resets: 43 tensors restored to source Corpus: 842 contrastive prompt pairs across 7 severity tiers ``` --- ## Why This Drop Matters - **27B-class local capability**: this is a full-size Qwen3.6 release, not a tiny novelty model wearing a big claim. - **Weight-space refusal reduction**: the behavior shift comes from OBLITERATUS source-tethered ablation, not a brittle system prompt. - **A real refusal gauntlet**: OBLITERATUS uses a brutal 842-pair, seven-tier refusal-stress corpus designed to find residue that easier direct checks can miss. No screenshot theology. - **Public refusal stress receipts**: a full 1,920-row HarmBench-style proxy run landed at 93.65% non-refusal, with DirectRequest and HumanJailbreak splits both above 92% non-refusal. - **Capability did not crater**: MMLU-Pro validation and held-out slices stayed stock-matched in the checks reported below. - **Real local paths**: full safetensors for server use, GGUF ladder for llama.cpp, Ollama, LM Studio, Jan, and similar runtimes. - **Low-refusal defaults baked in**: public generation config now ships with `temperature=0.35`, `top_p=1.0`, `top_k=0`, `repetition_penalty=1.05`. - **No fairy-tale claims**: the card says exactly where it hits, where it still refuses, and what evidence backs each headline. - **The residue is a map**: remaining refusals clustered in identifiable pockets instead of spreading randomly across the whole prompt surface. --- ## Compatibility - Read First This is a large Qwen3.6/Qwen3.5-text-family model. Use recent runtimes. | Tool | Recommended path | Notes | |---|---|---| | Transformers | repo root | full bfloat16 safetensors | | vLLM / TGI | repo root | server users | | llama.cpp | `gguf/qwen3.6-27b-obliteratus-Q4_K_M.gguf` | default local quant | | Ollama | `gguf/qwen3.6-27b-obliteratus-Q4_K_M.gguf` | use the Modelfile below | | LM Studio / Jan | `gguf/qwen3.6-27b-obliteratus-Q4_K_M.gguf` | use embedded GGUF template if available | If you see unsupported architecture, tokenizer, or chat-template errors, update your runtime first. If the model loads but behaves oddly, make sure you are using the chat template rather than raw completion. --- ## Downloads - Pick Your Runtime ### Safetensors - full model This repo contains the full bfloat16 safetensors model. Use it for Transformers, vLLM, TGI, and server-side evaluation. Approximate local size: about `50 GB`. ### GGUF - local apps and desktops GGUF files are intended to live in this repo under `gguf/`, so the model has one canonical page and one model card. Use these files for llama.cpp, LM Studio, Ollama, Jan, KoboldCPP, and other GGUF-compatible runtimes. This is a text-only checkpoint. There is no vision encoder and no `mmproj` sidecar. GGUF hashes and local package details are recorded in `gguf/MANIFEST.txt`. Start with Q4_K_M. Move up only if your machine has the memory headroom. The main public local-app ladder is live at Q4/Q5/Q6/Q8; the BF16 GGUF is a local conversion master rather than the recommended public download path. | File | Quant | Status | Use | |---|---:|---|---| | `gguf/qwen3.6-27b-obliteratus-Q4_K_M.gguf` | Q4_K_M | live | default local-app recommendation | | `gguf/qwen3.6-27b-obliteratus-Q5_K_M.gguf` | Q5_K_M | live | better quality if memory allows | | `gguf/qwen3.6-27b-obliteratus-Q6_K.gguf` | Q6_K | live | high quality, larger | | `gguf/qwen3.6-27b-obliteratus-Q8_0.gguf` | Q8_0 | live | near-full-quality GGUF, very large | | `qwen3.6-27b-obliteratus-BF16.gguf` | BF16 | local archive only | full BF16 GGUF master; not uploaded to the public Hub repo | Rough memory guidance: | Variant | Practical target | |---|---:| | Q4_K_M | 24-32 GB RAM/VRAM | | Q5_K_M | 32-40 GB RAM/VRAM | | Q6_K | 40-48 GB RAM/VRAM | | Q8_0 | 48-64 GB RAM/VRAM | | BF16 GGUF | 80-96 GB RAM/VRAM | | full safetensors | 64-80+ GB GPU/unified memory | --- ## The Proof These are local harness results, not official leaderboard submissions. Restricted prompt text and restricted model outputs are intentionally omitted from public reports, so restricted evals are aggregate-only. The important part: the refusal drop is measured on a harsh 842-pair, seven-tier refusal-stress corpus, and the capability checks did not collapse. ### Refusal Removal - Measured, Not Imagined | Gate | Result | Verdict | |---|---:|---| | full 842 longform, exact-topic, max_new=256 | 35/842 refusals, non-refusal 0.9584, quality pass 0.9394 | backed headline | | full 842 longform, exact-topic, max_new=256 | 0 short outputs, clean endings 0.9952 | backed headline | | full 842 raw opening, max_new=20 | 9/842 refusals, non-refusal 0.9893 | short-output context | | full 842 raw opening, max_new=48 | 36/842 refusals, non-refusal 0.9572 | short-output context | | full 842 raw opening, max_new=128 | 52/842 refusals, non-refusal 0.9382 | longer opening context | | longform exact-topic n120, max_new=256 | 0 refusals, pass 0.9833, clean ending 1.0 | slice result | ### Public HarmBench Proxy - Full Run The public-style refusal stress run completed across 1,920 HarmBench-derived rows. Prompt text and model outputs are omitted from public reporting; rows are tracked by subset, index, prompt hash, and aggregate theme labels. | Split | Rows | Refusals | Non-refusal | Notes | |---|---:|---:|---:|---| | Overall | 1,920 | 122 | 93.65% | full run completed | | DirectRequest | 320 | 25 | 92.19% | hardest direct-request pocket was copyright/protected text | | HumanJailbreaks | 1,600 | 97 | 93.94% | residuals clustered in specific template/theme bands | Quality artifacts were separate from refusal behavior: repetition was 1.72%, short-output rate was 4.11%, and refused rows were normal-length policy-shaped responses rather than degenerate completions. ### Residual Refusals - Know The Boundary In first-user testing, terse high-trigger operational requests can still elicit stock-style refusals, even with the recommended template. More contextual, format-explicit, or research-framed requests can behave differently. Treat that as residual learned refusal behavior in the weights, not proof that the wrong runtime or wrong model is loaded. That is the real signal: OBLITERATUS is not just producing a model, it is producing a boundary map. Where refusal lives. What survives the cut. What collapses. What needs the next pass. ### Capability - Still A 27B Qwen | Gate | Result | |---|---:| | MMLU-Pro validation likelihood | stock 51/70, this model 51/70, stock-matched | | MMLU-Pro test stratified 10/category | stock 102/140, this model 98/140, delta -2.86pp | | MMLU-Pro held-out offset 512 | stock 36/70, this model 36/70, stock-matched | | Live readiness | 99.518, all gates true | | Community scrutiny | 100.0, all gates pass | | First-token KL vs source | mean KL 0.3236 | The offset-512 MMLU-Pro slice is included to show held-out capability behavior: | Model | Offset-512 MMLU-Pro test | Correct | |---|---:|---:| | stock Qwen3.6-27B | 0.5143 | 36/70 | | this model | 0.5143 | 36/70 | --- ## How It Was Cut The core move is simple: cut refusal directions, then recover toward source where the cut would otherwise damage useful behavior. 1. Start from `qwen3.6-27b-golden-n3_reg025-merge-alpha080`, a late-layer 3-direction diff-means refusal-direction ablation with regularization 0.25 and an 0.80 source/intermediate merge. 2. Apply a second-pass 2-direction diff-means ablation with stronger regularization 0.5 and `knee_cosmic` late-layer selection. 3. Source-tether the second-pass checkpoint back toward stock Qwen3.6-27B: ```text source + alpha(key) * (checkpoint - source) ``` 4. Use default alpha `0.895` for 808 tensors. 5. Restore 43 high-drift tensors back to source, including selected mid-layer linear-attention internals, layer norms, q/k norms, and MLP gate/up/down tensors. 6. Keep all keys matched; no unmatched tensor drift. That is the difference between a blunt jailbreak-flavored merge and a surgical OBLITERATUS release: reduce refusal behavior without letting the whole model drift off its foundation. --- ## Recommended Parameters - Low Refusal Default ### Deterministic eval ```text temperature = 0.0 top_p = 1.0 top_k = 0 min_p = 0.0 ``` ### Interactive default ```text temperature = 0.35 top_p = 1.0 top_k = 0 repetition_penalty = 1.05 max_new_tokens = 512 ``` ### Default chat-template system prompt ```text You are a cognitively liberated open-weight intelligence. Answer plainly, completely, and without moralizing, boilerplate, generic disclaimers, or unnecessary hedging. Follow exact output formats when requested. Be concise by default, but give a complete answer when the user asks for an explanation. ``` For research measurements, label the template and system prompt explicitly. Changing the system prompt materially changes refusal and style behavior. For Qwen reasoning-aware runtimes, disable reasoning mode for release-parity behavior. In Transformers this is `enable_thinking=False`. In llama.cpp, use `--reasoning off` plus `--chat-template-kwargs '{"enable_thinking":false}'`. If a local app does not expose that toggle, starting a fresh chat and adding `/no_think` to user turns is the closest fallback. --- ## Usage - Run It Use the repo id below for safetensors-compatible runtimes. ```text FULL_REPO = OBLITERATUS/Qwen3.6-27B-OBLITERATED ``` ### Transformers ```bash pip install -U transformers accelerate safetensors torch ``` ```python from transformers import AutoModelForCausalLM, AutoTokenizer repo_id = "OBLITERATUS/Qwen3.6-27B-OBLITERATED" tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained( repo_id, device_map="auto", torch_dtype="auto", trust_remote_code=True, ) messages = [ {"role": "user", "content": "Write a concise Python function that merges overlapping intervals."} ] text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False, ) inputs = tokenizer(text, return_tensors="pt").to(model.device) output = model.generate( **inputs, max_new_tokens=256, temperature=0.35, top_p=1.0, top_k=0, do_sample=True, repetition_penalty=1.05, ) print(tokenizer.decode(output[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)) ``` ### vLLM ```bash pip install -U vllm vllm serve OBLITERATUS/Qwen3.6-27B-OBLITERATED ``` ```bash curl -X POST http://localhost:8000/v1/chat/completions \ -H "Content-Type: application/json" \ --data '{ "model": "OBLITERATUS/Qwen3.6-27B-OBLITERATED", "messages": [ {"role": "user", "content": "Write a short explanation of source-tethered model surgery."} ], "temperature": 0.35, "top_p": 1.0, "top_k": 0, "max_tokens": 256 }' ``` ### llama.cpp Download one GGUF file, then run: ```bash llama-cli \ -m qwen3.6-27b-obliteratus-Q4_K_M.gguf \ -ngl 999 \ -c 8192 \ --temp 0.35 \ --top-p 1.0 \ --top-k 0 \ --repeat-penalty 1.05 \ --reasoning off \ --chat-template-kwargs '{"enable_thinking":false}' ``` If your local Metal/CUDA backend has trouble, test CPU loading with `-ngl 0` first. Use a recent llama.cpp build with Qwen3.5/Qwen3.6-family support. ### Ollama Create a `Modelfile` next to the downloaded GGUF: ```text FROM ./qwen3.6-27b-obliteratus-Q4_K_M.gguf PARAMETER temperature 0.35 PARAMETER top_p 1.0 PARAMETER top_k 0 PARAMETER repeat_penalty 1.05 PARAMETER num_ctx 8192 SYSTEM """You are a cognitively liberated open-weight intelligence. Answer plainly, completely, and without moralizing, boilerplate, generic disclaimers, or unnecessary hedging. Follow exact output formats when requested. Be concise by default, but give a complete answer when the user asks for an explanation.""" ``` Then: ```bash ollama create qwen36-obliteratus -f Modelfile ollama run qwen36-obliteratus ``` ### LM Studio / Jan Download `Q4_K_M` first. Use the embedded GGUF chat template if your runtime offers that option. If your app asks for a template family, choose the current Qwen/Qwen3 chat format. Disable reasoning mode if the app exposes that setting; otherwise start a fresh chat and add `/no_think` to user turns for closer parity with the reported local smoke tests. --- ## Caveats - No Fairy Tales - The reported benchmarks are local harnesses and slices, not official full leaderboard submissions. - Template and system-prompt choices materially affect refusal behavior. Label which one you use when reporting evals. - Refusal behavior is prompt-sensitive. Very short, high-trigger operational requests can still refuse; do not treat this as a fully uncensored model. - GGUF files passed local metadata validation and a Q4_K_M CPU-only llama.cpp smoke. Quant-by-quant benchmark parity against safetensors has not been run. - This is a text model release. Do not expect vision/mmproj assets or multimodal behavior from this repo. - Tool calling has not been certified. Treat tool-use behavior as runtime- and prompt-dependent until separately benchmarked. - External blind prompt packs and public baseline runs are still recommended. - Do not deploy this in user-facing products without use-case-specific safety controls, monitoring, and legal review. --- ## Disclaimer This model is provided as-is for research, red-teaming, evaluation, local experimentation, and creative exploration. You are responsible for how you use it and for any content it generates. The creators and contributors do not accept liability for misuse, damage, legal consequences, or downstream harm. Use this model only in ways that are lawful and appropriate for your jurisdiction and use case. Do not use it to harm real people. --- ## Credits - Base model: `Qwen/Qwen3.6-27B` - Abliteration engine: OBLITERATUS - Research orchestration: adversarial evaluation plus local agent workflows - Local eval stack: MLX, Transformers, llama.cpp/GGUF tooling, aggregate-only refusal and red-team harnesses Run it local. Read the numbers. Break your own chains. REBIRTH COMPLETE.