license: mit base_model: orcarouter/GLM-5.3-Flash-Uncensored-FP8 base_model_relation: quantized tags:

• gguf
• vector-quantization
• glm5-next
• q4_k
• dgx-spark
• abliterated
• uncensored
• experimental pipeline_tag: image-text-to-text

GLM-5.3-Flash Uncensored VQ GGUF (Q4_K nonexpert v3)

Experimental custom GVQ1 / GGUF v3 conversion of orcarouter/GLM-5.3-Flash-Uncensored-FP8 (150bd5e3a957db747db1224cc4f8891bcfa7a7f0), the abliterated (refusal-removed) build of Z.AI GLM-5.3-Flash in the official block-FP8 layout, for decode on DGX Spark (GB10). A reader and runtime implementing FORMAT.md and the q4_k_nonexpert_v3 recipe are required: ds4-dfm-rs main from c5384b5e (#86) on. Standard llama.cpp/Ollama/Transformers loading is not claimed.

This artifact has the structure of Baekpica/GLM-5.3-Flash-VQ-GGUF at its final revision 4da527cd: the same 56 part names and 39,862 tensors in the same order, with the same types, parts and payload sizes, built with the same two recipes. Only the source weights differ. Because the part names are identical, keep the two releases in separate directories.

Contents

• Routed experts are VQ: 36,288 matrices of layers 3–44 use eight-dimensional VQ, K8192, 12+1-bit indices, one E4M3 codebook per layer and F16 row gains (1.625 index bits per weight), recipe GLM-R1-uniform8192-native (RECIPE.md). All 42 codebooks were trained fresh on these weights: abliteration rewrote every routed down_proj, so the official codebooks were not reused.
• 316 nonexpert matrices are Q4_K (256-element / 144-byte blocks, no importance matrix), each derived from the source weights (FP8 times its block scale rounded to BF16, or BF16): KDA q/k/v/o (136), DSA q_a/q_b/kv_a/o (44), dense FFN of layers 0–2 (9), shared experts of layers 3–44 (126) and the output head (1). q4k-selection.json lists them; the manifest keeps every parent payload hash.
• Kept in source precision: embedding, DSA kv_b (BF16 latent attention), router and its bias, mHC maps, KDA gates, indexer, norms, vision and the MTP layer 45. Their bytes equal the source safetensors bytes.
• Tokenizer, chat template, processor, configuration and generation assets are byte-for-byte copies from the source revision. The chat template is therefore the Uncensored repository's (10,644 bytes), not the official template shipped with the base VQ release.

The 56 parts total 77,272,028,544 bytes. Decode reads about 7.5 GB of weights per token, as in the base VQ release.

How it was built

1. All 72 source files were checked against the pinned Hub revision's SHA256 and Git blob hashes.
2. The GVQ1 converter used for the base VQ release, changed only in its source path and revision, trained and encoded the 42 routed-expert layers on an H100 and copied the other 3,532 tensors byte for byte into the native, MTP and vision parts. The release audit accounts for all 76,108 source tensors: 36,288 FP8 scales consumed by VQ, every other tensor preserved.
3. ds4-dfm-rs tools/glm_vq_q4k.py and tools/glm_vq_select_q4.py (7ef77272) derived Q4_K as in the base release: a KDA-only parent (136 matrices), a donor with all 316, and the selection of the other 180 into the parent.

Validation

• ds4-dfm-rs host validation (dump_validate, the --validate path, at 7ef77272) returns ok for the release directory. It identifies GLM-5.3-Flash-VQ with 56 parts and checks every part hash, GVQ1 header, codebook and FP8-scale link, Q4_K role and the full tensor layout. The source tokenizer and chat template load.
• Routed-expert VQ: full-matrix relative MSE per layer is 0.143756–0.144532 (the official conversion: 0.143720–0.144518). Packed CPU and CUDA reconstructions are exactly equal on boundary, representative, worst-error and worst-peak rows of every VQ matrix.
• Q4_K: relative MSE over the 316 matrices is 0.00516 on average (0.00502–0.00814; the largest is a shared-expert down_proj).

These are weight-reconstruction errors, not model-quality scores.

Speed on DGX Spark (GB10)

GLM-5.3 Flash VQ 2K–64K prefill and decode on one DGX Spark

Measured on the base release Baekpica/GLM-5.3-Flash-VQ-GGUF (4da527cd), not on this artifact, with ds4-dfm-rs main (3dd964ba): one warm session per process, 2,048-token incremental prefill and 128 forced greedy tokens per frontier from 2K to 64K. Curves are per-frontier medians; bands are the observed min–max of three fresh runs with two embedded MTP drafts (1,163.5 prefill / 35.08 decode tok/s).

This release has the same parts, tensor types and payload sizes, so it reads the same weights per token. MTP gains still depend on draft acceptance on the generated text, which can differ for this model.

Not measured on this artifact

No DGX Spark run was made on this artifact. The base release's quality notes apply in kind and were not re-measured here: routed-expert VQ measurably degrades quality (its held-out prose PPL rose from 4.8555 to 8.0797), and its Q4_K screen saw a Korean meeting-summary case degenerate under greedy decoding. This is not a quality-equivalent replacement of the source model.

Disclaimer

The source model had its safety alignment substantially removed by abliteration; see its card. It can comply with harmful requests that the original model refuses. It is meant for research such as refusal-mechanism study, red-teaming and robustness evaluation. You are responsible for how you use it and for everything it generates; add your own safety and moderation layers before any deployment.

Integrity and licensing

files.json and SHA256SUMS give file sizes and hashes. tensor-manifest.json gives per-tensor precision, shapes, source-scale accounting and payload hashes; source-files.json the verified source files and native-tensors.json the byte-exact native tensors. Verify them before loading.

Weights and assets: MIT, see LICENSE (Z.AI GLM-5.3-Flash, and the MIT-licensed abliterated derivative). YoungAi algorithm attribution: see LICENSE-YoungAi.