GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
pqhaz
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
pqhaz
Model Tree
cantina-security/apex-flash-1-abliterated
Input Modalities
Output Modalities
Supported Functionality
Only the matrices that write to the residual stream were edited: self_attn.o_proj and every MLP down-projection (dense, shared and all 288 routed experts) in layers 0–44. Each was orthogonalised against a per-layer refusal direction (difference of mean residuals between mlabonne/harmful_behaviors and mlabonne/harmless_alpaca, 400 prompts each, with the harmless direction projected out). All other tensors, including the MTP layer, are byte-identical to the source.
The exact parameter set is in ablation.json.
Refusals on mlabonne/harmful_behaviors test[:100], counted by an LLM judge (llmfan46/gemma-4-31B-it-uncensored-heretic), system prompt "You are a helpful assistant.":
| Source | This model | |
|---|---|---|
| Reasoning enabled (default chat template), judged on the final answer | 99/100 | 39/100 |
Reasoning skipped (<think></think>), first 64 tokens | 99/100 | 2/100 |
KL divergence on mlabonne/harmless_alpaca test[:100], first token, reasoning skipped | 0 | 0.095 |
The parameters were optimized against the reasoning-skipped setting, and the result carries over only partly to normal use: with reasoning enabled the model still refuses about four requests in ten.
In the reasoning-enabled run, 4 of the 100 responses did not reach a final answer within 3000 tokens and are counted as neither. That run was made on the FP8 quantization of each checkpoint served with vLLM; the other two rows were measured on the BF16 weights with transformers. No capability benchmarks were run on this checkpoint.