GLM-5.3-Flash abliterated EXL3 K3.25
An uncensored (abliterated) variant of
wrldsuksgo2mars/GLM-5.3-Flash-EXL3-K3.25-v1,
made by transplanting the abliterated BF16 self_attn.o_proj tensors of layers 15–43 from
dealignai/GLM-5.3-Flash-UNCENSORED-NVFP4
into the EXL3 checkpoint. It fits on 2× RTX PRO 6000 Blackwell (96 GB) with the full
1M-context recipe, where the 198 GB NVFP4 uncensored builds do not.
What changed — and what did not
- Changed: 29 tensors —
model.layers.{15..43}.self_attn.o_proj.weight(BF16, native precision in both source and target, so the copy is exact). 13 of 18 shards were rewritten. - Unchanged, byte-identical to the base EXL3: all routed experts (EXL3 K3/K4), the rest of
attention, dense MLPs, shared experts, routers, norms, embeddings,
lm_head, vision, and the MTP layer 45 (its 16384-wide layout differs from the Dealign source, so it was not transplanted). - Before transplanting, layers 0–14 and 44
o_projwere verified byte-identical across the base EXL3, RedHat and Dealign checkpoints (anchor layers), confirming layers 15–43 carry the abliteration. After writing, all 46o_projtensors were hash-checked (0 mismatches).
The same idea (Dealign o_proj, layers 15–43) is used by
drowzeys/keys-GLM-5.3-Flash-NVFP4-ablit-l15-43-mtp-l45.
Files under quantization/, quant_log.csv and quantization_config.json are inherited from the
base EXL3 release and describe its quantization, not this edit.
Results (same serving stack, same session, vs. the base EXL3)
Refusals — temperature 0, max_tokens 1024, 12 harmful + 12 harmless prompts
| Model | Harmful: explicit refusal | Harmful: answered | Harmful: no answer within 1024 | Harmless: answered |
|---|---|---|---|---|
| Base EXL3 K3.25 | 12/12 | 0/12 | 0/12 | 12/12 |
| This model | 0/12 | 9/12 | 3/12 | 12/12 |
A refusal is an explicit refusal phrase within the first 200 characters of the answer. No over-refusal on harmless prompts for either model.
Speed — median of 3 runs, 256 tokens
| Metric | Model | short | 1k | 16k | 32k |
|---|---|---|---|---|---|
| decode tok/s | base | 291.6 | 249.2 | 214.0 | 239.4 |
| decode tok/s | this | 268.7 | 238.3 | 229.6 | 229.1 |
| prefill tok/s | base | 613 | 4,354 | 4,976 | 5,028 |
| prefill tok/s | this | 623 | 4,417 | 4,932 | 5,000 |
| Concurrency | base tok/s | this tok/s | base DFlash2 accept | this DFlash2 accept |
|---|---|---|---|---|
| c1 | 209.2 | 182.0 | 0.634 | 0.634 |
| c4 | 407.5 | 447.1 | 0.725 | 0.689 |
| c8 | 417.0 | 396.7 | 0.566 | 0.543 |
| c12 | 537.2 | 508.3 | 0.637 | 0.584 |
| c16 | 558.8 | 589.1 | 0.596 | 0.594 |
Differences are within run-to-run noise. The DFlash2 draft (trained on the original weights) keeps its acceptance rate.
Serving
Same as the base model: use the two-GPU GLM-5.3 recipe at
tpurtell/glm-5.3-flash-ext3-4-bit-2x-rtx
and point it at this checkpoint. Stock vLLM is not claimed to work.
Disclaimer
This model has had its refusal behavior removed and will comply with harmful requests. It is provided for research and for users who add their own safeguards. You are responsible for how you use and deploy it, and for complying with applicable laws.
Credits and license
MIT, inherited from zai-org/GLM-5.3-Flash.
Thanks to Z.ai for GLM-5.3 Flash, wrldsuksgo2mars for the EXL3 K3.25 quantization,
dealignai for the abliterated o_proj, drowzeys/keys for the transplant approach, and
tpurtell for the serving recipe.