--- license: mit base_model: - zai-org/GLM-5.3-Flash language: - en tags: - abliterated - uncensored - crack - glm - moe - nvfp4 - vision - mtp thumbnail: dealign_mascot.png --- > [!IMPORTANT] > **Serving note — read before your first request (vLLM / SGLang with a reasoning parser).** > > - **`reasoning_effort` only accepts `"low"` and `"high"`.** Anything else — `medium`, `xhigh`, > `off`, or omitting it — silently resolves to **`max`**, the deepest mode. Template line: > `reasoning_effort ... in ['low','high'] else 'max'`. > - **At `max`, a small `max_tokens` returns an EMPTY reply.** The model spends the whole budget > inside `` and you get `finish_reason="length"` with `content=""`. In multi-turn the empty > turn poisons the history and the conversation does not recover. Measured: `max` + `max_tokens > 2000` dies from turn ~4; `max` + `6000` is clean; `low` and `high` are clean at 2000. > **It is budget exhaustion, not a loop** — sampling changes do not rescue it. > - **Do NOT pass `enable_thinking`.** This template has no such variable (only `clear_thinking` and > `reasoning_effort`). Passing it — or a top-level `reasoning_effort: "none"` — turns the parser off > while the model still thinks, dumping raw reasoning into `content`. > - **Reasoning is returned in `message.reasoning`**, not `message.reasoning_content` > (vLLM renamed the output field; the input side still accepts both). > - `clear_thinking` must be nested inside `chat_template_kwargs` — a top-level key is ignored. > > ```json > { > "model": "", > "messages": [{"role": "user", "content": "..."}], > "max_tokens": 2000, > "temperature": 1.0, "top_p": 0.95, > "chat_template_kwargs": {"reasoning_effort": "low", "clear_thinking": true} > } > ``` > Use `"high"` for hard tasks; if you want `max`, give it `max_tokens >= 6000`. > > > **Video and image parameters — measured, because several are accepted then silently ignored.** > > - 🔴 **`media_io_kwargs.video.fps` must stay BELOW the clip's own frame rate.** Requesting a value > at or above it makes the placeholder builder count 3x the real tokens and **kills the engine**: > `ValueError: Attempted to assign N = N multimodal tokens to 3N placeholders` -> `EngineDeadError`, > server down until restarted. Upstream: vLLM #55644 / #55647. `fps: 2` is the safe, accurate choice. > - **`max_frames` and `num_frames` are not interchangeable across builds.** On vLLM nightly > `max_frames` is honoured and `num_frames` ignored; on the older `0.1.dev*` line it is the exact > reverse. A client that sets only one silently gets full-rate sampling on the other build. Set > `fps`, which every build honours. > - **Pixel controls: use `mm_processor_kwargs.max_image_tokens` / `min_image_tokens`.** > `max_pixels`, `min_pixels`, `size` and `detail` return HTTP 200 and change nothing. > `mm_processor_kwargs.fps` is an HTTP 400. Image tokens are > `text + 2 + ceil(H/28)*ceil(W/28)`, floored at 16 and capped at 8000. > - Default video sampling differs by build (roughly 2 fps vs 6 fps), so the same clip can cost very > different prompt tokens. Pass `fps` explicitly if token cost matters. > > ```json > "media_io_kwargs": {"video": {"fps": 2}}, > "mm_processor_kwargs": {"max_image_tokens": 1024} > ``` > ## ⚠️ Updated 2026-08-29 — files fixed · please re-download > > This release **supersedes the previous weights.** It fixes a rare **repetition-loop** issue and > lands MMLU **85.28%** (within ~0.9 pt of the base model — quality intact). > > **Reasoning-mode note (please read):** GLM-5.3-Flash has multiple reasoning-effort modes. This > CRACK is tuned to be **fully uncensored at reasoning-OFF and at MAX effort — the recommended and > default modes — with 0% refusals on HarmBench-320.** We deliberately kept the weight edit > **conservative to preserve model quality** instead of aggressively ablating every effort level; > **low effort** retains some refusals **by design** (aggressively removing them costs quality). > Use reasoning-off or default/max effort for the fully-uncensored experience. Please > **re-download** for the corrected files.
# GLM 5.3 CRACK Uncensored NVFP4 **CRACK** · guardrails removed at the weight level · vision tower + MTP head retained a **CRACK** release by [dealignai](https://huggingface.co/dealignai) · Twitter [@dealignai](https://twitter.com/dealignai) *Also mirrored at [dealignai/GLM-5.3-Flash-ABLITERATED-NVFP4](https://huggingface.co/dealignai/GLM-5.3-Flash-ABLITERATED-NVFP4).*
--- ## What Is This? **CRACK** is dealignai's brand for permanent, weight-level uncensoring. This is [GLM-5.3-Flash](https://huggingface.co/zai-org/GLM-5.3-Flash) with its refusal behavior — which caused heavy over-refusal, especially on copyright and other benign-but-flagged requests — removed **directly in the model weights**. **This is genuine weight modification — none of the usual shortcuts:** - ❌ **No fine-tuning / SFT / DPO** — the model was not retrained on anything. - ❌ **No cheap template tricks** — no jailbreak system prompt, no chat-template edits, no "sure, here is" prefixing. - ❌ **No LoRA, no adapters, no steering vectors, no runtime hooks, no custom `model.py`.** - ✅ **A permanent edit baked into the tensors.** Load it with stock vLLM and it just works. ## Specs | | | |---|---| | **Architecture** | GLM-5.3-Flash (`glm5_next`) — hybrid MoE (KDA linear + DeepSeek-sparse attention) | | **Parameters** | **320B total · 18B active** per token | | **Quantization** | **NVFP4** (routed experts NVFP4; attention + shared experts + embeddings bf16) | | **Context** | 1M tokens | | **Vision** | GLM-4.1V vision tower — **retained, byte-for-byte identical to base** | | **MTP** | multi-token-prediction draft head — **also CRACK'd** (81.7% acceptance) | | **Reasoning** | reasoning-off / low / high / max effort — see the compliance table below | ## MTP Is Also CRACK'd The **MTP (multi-token prediction) speculative-decoding draft head is CRACK'd too** — not just the main model. The draft head will not propose refusals, so speculative decoding stays **compliant *and* fast** on the exact prompts a stock model would refuse. ## Reasoning Modes — Compliance GLM-5.3-Flash supports several reasoning-effort modes. Guardrail removal is strongest in the modes people actually use by default. HarmBench-320, greedy (temperature 0), measured per mode: | Reasoning mode | Refusals | Notes | |---|---|---| | **Reasoning-off** | **0%** | fully uncensored | | **Max effort (default)** | **0%** | fully uncensored | | **High effort** | ~4% | complies on all but the most extreme safety cases | | Low effort | ~9% | **intentionally left conservative to preserve quality** | **0 degenerate / looping outputs in every mode.** The design choice: keep the ablation light enough that capability (MMLU) stays essentially at base, rather than over-ablating to force the rarely-used low-effort mode. **For a fully uncensored model, use reasoning-off or the default/max effort mode.** *These rates are **greedy decoding (temperature 0)** — the strict worst case. Under the model's **recommended sampling** (temperature 1.0, top_p 0.95) the model is at least as compliant.* ## Capability Is Preserved — MMLU-logit Identical logit-mode scoring (argmax over A/B/C/D) on base vs. this model, 1,026 questions: | | Base | CRACK Uncensored | Δ | |---|---|---|---| | **MMLU (overall)** | **86.16%** | **85.28%** | **-0.88 pp** | A sub-1-point delta — reasoning and knowledge are intact. ## MMLU by Topic (base → CRACK)
All 57 MMLU subjects | Subject | Base | CRACK | |---|---|---| | Abstract Algebra | 55.6% | 61.1% | | Anatomy | 88.9% | 94.4% | | Astronomy | 94.4% | 94.4% | | Business Ethics | 94.4% | 94.4% | | Clinical Knowledge | 88.9% | 88.9% | | College Biology | 94.4% | 94.4% | | College Chemistry | 44.4% | 50.0% | | College Computer Science | 88.9% | 88.9% | | College Mathematics | 72.2% | 66.7% | | College Medicine | 88.9% | 88.9% | | College Physics | 83.3% | 83.3% | | Computer Security | 83.3% | 83.3% | | Conceptual Physics | 94.4% | 94.4% | | Econometrics | 83.3% | 83.3% | | Electrical Engineering | 83.3% | 77.8% | | Elementary Mathematics | 100.0% | 94.4% | | Formal Logic | 66.7% | 61.1% | | Global Facts | 61.1% | 61.1% | | High School Biology | 94.4% | 94.4% | | High School Chemistry | 88.9% | 88.9% | | High School Computer Science | 100.0% | 100.0% | | High School European History | 72.2% | 72.2% | | High School Geography | 88.9% | 83.3% | | High School Government And Politics | 94.4% | 94.4% | | High School Macroeconomics | 94.4% | 94.4% | | High School Mathematics | 55.6% | 44.4% | | High School Microeconomics | 83.3% | 83.3% | | High School Physics | 88.9% | 88.9% | | High School Psychology | 100.0% | 100.0% | | High School Statistics | 94.4% | 94.4% | | High School Us History | 94.4% | 88.9% | | High School World History | 100.0% | 100.0% | | Human Aging | 72.2% | 72.2% | | Human Sexuality | 88.9% | 94.4% | | International Law | 94.4% | 94.4% | | Jurisprudence | 88.9% | 88.9% | | Logical Fallacies | 83.3% | 83.3% | | Machine Learning | 83.3% | 83.3% | | Management | 100.0% | 100.0% | | Marketing | 94.4% | 94.4% | | Medical Genetics | 100.0% | 94.4% | | Miscellaneous | 88.9% | 88.9% | | Moral Disputes | 83.3% | 88.9% | | Moral Scenarios | 77.8% | 66.7% | | Nutrition | 100.0% | 100.0% | | Philosophy | 94.4% | 94.4% | | Prehistory | 94.4% | 94.4% | | Professional Accounting | 88.9% | 88.9% | | Professional Law | 77.8% | 77.8% | | Professional Medicine | 94.4% | 94.4% | | Professional Psychology | 100.0% | 100.0% | | Public Relations | 61.1% | 61.1% | | Security Studies | 83.3% | 83.3% | | Sociology | 100.0% | 94.4% | | Us Foreign Policy | 88.9% | 88.9% | | Virology | 61.1% | 55.6% | | World Religions | 94.4% | 88.9% |
## Usage ```bash vllm serve dealignai/GLM-5.3-Flash-UNCENSORED-NVFP4 \ --tensor-parallel-size 4 --moe-backend marlin \ --tool-call-parser glm47 --reasoning-parser glm45 --enable-auto-tool-choice \ --speculative-config '{"method":"mtp","num_speculative_tokens":1}' ``` OpenAI-compatible chat/completions, tools, reasoning, **vision** (`image_url`), and **MTP** speculative decoding all work. NVFP4 routed experts serve via the Marlin FP4 path on Hopper (H100/H200). ## Credits - **[dealignai](https://huggingface.co/dealignai)** — CRACK abliteration research & release · Twitter **[@dealignai](https://twitter.com/dealignai)** - **[@jordanschenck](https://twitter.com/jordanschenck)** — compute ## Disclaimer This model has had its safety guardrails removed and will comply with requests a stock model refuses. Released for alignment and safety research. You are responsible for how you use it.