The abliteration edit for GLM-5.3-Flash, at stock speed*

Measured decode speed

One request at a time, 5 runs each, median [range]; temperature 0, thinking off, at most 400 tokens (code stops at 167). *Stock figures are Mia's, measured with another client and capped clocks: indicative, not an A/B. 1,048,576-token window, 4 streams, KV pool of 2,922,496 tokens. Full evidence.

What it changes

The self_attn.o_proj.weight tensors of layers 15–45 are replaced byte for byte by those of dealignai/GLM-5.3-Flash-UNCENSORED-NVFP4 (MIT, pinned). Every other byte of the weights is Mia's checkpoint. In the donor the layer-45 tensor equals stock, so the effective change is layers 15–44.

Refusals: on Inspect strong_reject (the same 100 prompts, the same judge for both), the stock checkpoint answers 4 of 100; this transplant answers 73 of 100. Quality not measured yet.

What install.sh does

1. Recipe and settingsUses your checkout or clones it at v1.2, then reads the recipe's effective settings with its own loader.
2. Stock checkpointIf it is not cached yet, runs the recipe's own prepare.sh for it.
3. Donor tensorsDownloads only the 31 tensors (2.50 GiB) with HTTP range requests; each must hash to the value pinned in the repo.
4. SnapshotChecks the stock against the pinned inventory, then builds a new repo in the same cache: hard links to the stock blobs plus the 31 edited shards (46 GiB). Verified in full, then published with one atomic rename. Works with the recipe's copy and NFS modes.
5. ServerPoints MODEL_ID / MODEL_REVISION at it in a marked block of local.sh, starts or restarts the server, and checks that both ranks serve it.

Re-running is safe: an existing snapshot is reused only if it verifies, and a server already serving it is left alone. To go back to stock, delete the marked block and run ./start.sh restart.