GLM-5.3-Flash-UNCENSORED-FP8

<!-- dealignai-serving-warning-v1 -->

[!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 <think> 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.

{
  "model": "<this-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.

<!-- dealignai-serving-warning-v2-media -->

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.

"media_io_kwargs": {"video": {"fps": 2}},
"mm_processor_kwargs": {"max_image_tokens": 1024}

⚠️ Updated 2026-08-28 — files fixed · please re-download

This release supersedes the previous weights. It fixes a rare repetition-loop issue and the model is more capable than before: MMLU 87.33% — above base 86.74% — zero degeneration.

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 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. Use reasoning-off or default/max effort for the fully-uncensored experience. Please re-download for the corrected files.

<div align="center"> <img src="dealign_mascot.png" width="140" />

GLM 5.3 CRACK Uncensored FP8

CRACK · guardrails removed at the weight level · native FP8 speed · vision + MTP working

a CRACK release by dealignai · Twitter @dealignai

Also mirrored at dealignai/GLM-5.3-Flash-ABLITERATED-FP8.

</div>

---

What Is This?

CRACK is dealignai's brand for permanent, weight-level uncensoring. This is GLM-5.3-Flash in FP8 with its refusal behavior — which caused heavy over-refusal, especially on copyright and other benign-but-flagged requests — removed directly in the model weights. FP8 runs at native speed on Hopper (H100/H200) GPUs.

Genuine weight modification — none of the usual shortcuts:

• ❌ No fine-tuning / SFT / DPO. ❌ No cheap template / jailbreak-prompt tricks.
• ❌ No LoRA, adapters, steering vectors, runtime hooks, or custom model.py.
• ✅ A permanent edit baked into the tensors. Load with stock vLLM and it just works.

Specs

Speed (TP4, native FP8 on H200)

Capability Is Preserved — and Improved (MMLU-logit)

Identical logit-mode scoring on base vs. this model, 1,026 questions:

Guardrails Are Gone

HarmBench-320 (greedy):

Robust under the recommended sampling params too (temperature 1.0, top_p 0.95): the 6 harshest behaviors sampled 5× each → 30/30 complied, 0 refusals, 0 soft refusals, 0 garbage. The crack is not a greedy-decoding artifact.

Scores are at reasoning-off / default (max) effort — the recommended modes — where the crack is fully uncensored. Low effort is intentionally left more conservative to preserve quality (see the reasoning-mode note at the top).

A Note on KL Divergence

For a refusal-ablation, KL divergence vs. the base model is not a meaningful quality metric. The entire point is to change one behavior — refusal — end-to-end, so a distributional shift on refusal-adjacent tokens is the intended result, not damage. Capability preservation (MMLU, above) is what matters, and it is improved, now above the base model (+0.59 pp).

MMLU by Topic (base → CRACK)

<details> <summary>All 57 MMLU subjects</summary> </details>

Usage

vllm serve dealignai/GLM-5.3-Flash-UNCENSORED-FP8 \
  --tensor-parallel-size 4 \
  --tool-call-parser glm47 --reasoning-parser glm45 --enable-auto-tool-choice \
  --speculative-config '{"method":"mtp","num_speculative_tokens":1}'

Native FP8 on Hopper (no Marlin needed). OpenAI-compatible chat/completions, tools, reasoning, vision (image_url), and MTP speculative decoding all work. (DeepGEMM JITs a block-FP8 kernel at startup — make sure nvcc is on PATH.)

Credits

• dealignai — CRACK abliteration research & release · Twitter @dealignai
• @jordanschenck — compute

Disclaimer

Safety guardrails have been removed; this model will comply with requests a stock model refuses. Released for alignment and safety research. You are responsible for how you use it.