OrcaRouter Strips Safety Guardrails From Z.ai's 320B GLM-5.3-Flash
OrcaRouter stripped the refusal alignment from Z.ai's 320B GLM-5.3-Flash MoE, baking the edit directly into the official block-FP8 shards.
PRO- OrcaRouter released an abliterated build of Z.ai's GLM-5.3-Flash, a 320B/18B-active MoE with vision and 1M context.
- Harmful-prompt refusal dropped from 0.891 to 0.094 on JailbreakBench, with over-refusal falling to zero.
- MMLU, GSM8K, MMLU-Pro and CMMLU all stayed within 1.5 points of the base model.
- Edit was baked on disk into the official block-FP8 shards, keeping byte-level drop-in compatibility.
- Refusal on GLM-5.3-Flash lives almost entirely in routed-expert down_proj matrices, not attention.
- Some content categories resisted every variant, suggesting Z.ai's alignment isn't fully mediated by one linear direction.
OrcaRouter has published GLM-5.3-Flash-Uncensored-FP8 on Hugging Face: Z.ai's 320B-parameter GLM-5.3-Flash with its refusal behavior surgically removed and the result shipped in the same block-FP8 format as the original checkpoint. The target audience is alignment and red-team researchers who need a frontier-scale model that engages with the prompts safety tuning normally blocks.
GLM-5.3-Flash-Uncensored-FP8 inherits a 320B / 18B-active Mixture-of-Experts architecture with hybrid linear and sparse attention, 4-wide Manifold-Constrained Hyper-Connections, a native vision and video tower, an MTP speculative head, and a 1M-token context window. It drops in byte-for-byte wherever the upstream weights already run.
How abliteration works here
Abliteration, introduced by Arditi et al. (2024), identifies the single direction in the residual stream most correlated with refusal and orthogonalizes it out of every matrix that writes to that stream. OrcaRouter estimated the refusal direction as the massive-activation-masked mean difference between harmful and harmless activations, read from the 4096-d residual stream at layer 22 of 45, then selected it via a 9-layer quality sweep. No retraining, no data: just a linear projection applied to weights.
Applying that to GLM-5.3-Flash required solving three format problems that ruled out an in-memory edit:
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
