Skip to main content
Back

dealignai Rewrites GLM-5.3-Flash Weights to Strip All Refusals

A weight-level abliteration of Z.ai's 320B MoE strips refusals while nudging MMLU up 0.59 points, with vision and speculative decoding intact.

  • dealignai released GLM-5.3-Flash-UNCENSORED-FP8, a weight-level abliteration of Z.ai's 320B MoE
  • Scores 87.33% on MMLU, 0.59 points above the FP8 base model, with no fine-tuning used
  • HarmBench-320 compliance hits 100% at reasoning-off and max effort modes
  • Runs at 163 tok/s on H200 TP4, 211 tok/s with MTP speculative decoding enabled
  • Keeps vision tower, 1M-token context, and MIT license from the base GLM-5.3-Flash
  • Low reasoning effort intentionally left partially censored to preserve capability on benchmarks

An organization using the name dealignai has published an FP8 version of Z.ai’s base checkpoint with refusal behavior edited directly in the weights. The group labels the edited checkpoint “uncensored” and reports a small MMLU gain after the modification. The benchmark and performance results come from the publisher and have not been independently reproduced.

Release at a glance
Field Details
Model size About 321 billion total parameters, with 18 billion active per token
Weight format Block-wise e4m3 FP8
Context limit Up to 1 million tokens
Modification Direct tensor edit intended to reduce refusals
Serving Stock vLLM without custom model code or runtime hooks
License MIT, according to the edited repository

Direct weight edit, intact architecture

GLM-5.3-Flash is a multimodal mixture-of-experts model, which routes each token through a subset of its 321 billion parameters to reduce inference cost. Z.ai describes it as the first GLM model to combine linear and sparse attention in one stack. Its 45 attention layers comprise 34 Kimi Delta Attention layers and 11 DeepSeek Sparse Attention layers, interleaved in a 3:1 pattern.

The edited release retains the vision tower, which handles image inputs, and the multi-token-prediction head used for speculative decoding. Its block-wise e4m3 format stores weights as 8-bit floating-point values with shared scales across blocks, allowing Hopper-generation tensor cores to execute the relevant operations efficiently.

The repository says the modification required none of the usual runtime or adapter machinery:

  • No supervised fine-tuning or preference optimization
  • No LoRA adapters, steering vectors, or inference hooks
  • No custom model.py implementation
  • No jailbreak prompt or chat-template dependency

Embedding the change in the checkpoint makes it persist across reloads and standard vLLM deployments. Refusal suppression still varies by reasoning mode, so the “uncensored” label does not describe every configuration.

A narrow benchmark gain needs context

MMLU in logit mode evaluates multiple-choice questions by comparing the probabilities assigned to answer options. On the publisher’s 1,026-question run, the edited model scored 87.33%, compared with 86.74% for the FP8 base.