4× DGX Spark TP4 — our EXL3 quant of the uncensored GLM-5.3-Flash — August 29, 2026
176 GB EXL3 TR3 4bpw · 1M context · ~97 tok/s single-stream · DFlash2 k=7 spec decode · parity with aligned base
Encoded on 4 DGX Sparks in 4h22m (37,152 expert tensors) from orcarouter uncensored FP8 weights. Abliteration survives quantization — verified by discriminator probe.
I only post models I actually use. — neko-legends
# GLM-5.3-Flash Uncensored — EXL3 TR3 4bpw
## Licensing and attribution
> This work includes or was produced using ShapleyMcg, created by Brandon M. Music
> (https://github.com/brandonmmusic-max/shapleymcg). ShapleyMcg is licensed under the
> ShapleyMcg License v1.0, an attribution-required license that grants no rights to
> the person known as "0xSero." Use of ShapleyMcg without this attribution is unlicensed.
This checkpoint is a mixed-license artifact:
- **Model weights chain** — `zai-org/GLM-5.3-Flash` (MIT) → `orcarouter` uncensored
FP8 derivative (MIT). The uncensored weights are redistributable under MIT terms.
- **Calibration artifacts** — the per-expert `suh`/`svh`/`mcg` scale tensors were
inherited from the ShapleyMcg-calibrated checkpoint
([brandonmusic/GLM-5.3-Flash-tr3-4bpw](https://huggingface.co/brandonmusic/GLM-5.3-Flash-tr3-4bpw),
mirrored by Mia-AiLab), and the trellis codes were produced by running the published
R10 encoder closure. Those portions are licensed under the
[ShapleyMcg License v1.0](https://github.com/brandonmmusic-max/shapleymcg/blob/main/LICENSE),
included in this repository as `ShapleyMcg-LICENSE`.
- **Our contributions** (the uncensored-weight re-encode, non-expert dequantization,
verification, benchmarks, this card) may be used under either license above.
```bibtex
@misc{music2026shapleymcg,
author = {Music, Brandon M.},
title = {ShapleyMCG: An Auditable Calibration-to-Encoding Pipeline for
Low-Bit Mixture-of-Experts Models},
year = {2026},
url = {https://github.com/brandonmmusic-max/shapleymcg},
note = {Licensed under the ShapleyMcg License v1.0}
}
```
## Intended use and status
This is an **uncensored (abliterated) model**: upstream refusal alignment has been
removed from the weights. It exists for the work where that matters — and that work
is overwhelmingly legitimate:
- **Defending your own systems.** Auditing, hardening, and penetration-testing
infrastructure you own or are authorized to test — your home lab, homelab network,
self-hosted services, and the software you run. Real adversarial testing needs a
model that does not refuse to think like an attacker.
- **Red-team and purple-team exercises** under authorization, security research,
adversarial evaluation, and defensive analysis of malicious content (phishing
triage, malware analysis, social-engineering resistance training).
- **Self-hosted deployment** where a household or individual wants a model that
behaves like a capable colleague, not a compliance department — on their own
hardware, for their own purposes.
**Please use responsibly.** This model will help with things it should not be
used for. The operators of this repository publish it for authorized security
research and personal self-hosted use; users are responsible for complying with
applicable laws and the rules of any system they point it at. Do not use it
against systems you do not own or lack explicit permission to test.
No capability, safety, or bias evaluations beyond the benchmarks in this card
have been performed. Access is gated; by requesting access you accept these terms.
## Format
Same format as [Mia-AiLab/GLM-5.3-Flash-EXL3-TR3-4bpw](https://huggingface.co/Mia-AiLab/GLM-5.3-Flash-EXL3-TR3-4bpw)
(`quantization_config`: bits 4, codebook mcg, scope `glm53_routed_experts_only`, exllamav3 0.0.43):
- **Routed experts** (288 × 42 MoE layers): EXL3/K4 trellis codes + fp16 `suh`/`svh` block scales
- **Everything else** (attention, NoPE MLA, shared experts, dense layers, embeddings): BF16,
dequantized from the FP8 source (official-source-native policy)
- `lm_head` stays fp16/bf16
### Quantization recipe
The 4-bit encode was produced with brandonmusic's published R10 encoder closure
(`reproducibility/r10` from [brandonmusic/GLM-5.3-Flash-tr3-4bpw](https://huggingface.co/brandonmusic/GLM-5.3-Flash-tr3-4bpw)):
- **Scale vectors (`suh`/`svh`/`mcg`) inherited per-tensor from the Mia-AiLab base EXL3
checkpoint** — reusing its calibrated scale search (the uncensored weights are a small
perturbation of base)
- Trellis codes re-encoded from the FP8-dequantized uncensored weights
- Identity covariance (no activation capture this run); pilot round-trip rel-err ≈ 6.75%
per expert tensor
- Encoded in parallel across 4 Sparks: 37,152 tensors in 4h22m
## Serving
Use the MiaAI-Lab EXL3 kit (`start.sh` / our TP4 variant) with the same settings as the
aligned checkpoint. Verified on 4× DGX Spark TP4:
| Metric | This quant | Base EXL3 (same protocol) |
|---|---|---|
| C1 decode, usage-counted: code / structured / math / prose | 86 / 113 / 69 / 37 tok/s | 83 / 113 / 71 / 33 tok/s |
| Single-stream ground truth (450 tok, wall clock) | 97 tok/s | 97 tok/s |
| C4 concurrent aggregate (4× identical) | 220 tok/s | 150 tok/s (warmer cache favors this run; treat as parity) |
| KV pool @ 1M context | 6.13M tokens | 6.14M tokens |
| DFlash2 spec decode (k=7) | works; accept ~48% avg on reasoning prose vs ~84% base (drafter trained on base hidden states) — wall-clock speed parity | ~84% |
| Cold prefill ~380k window (pre-E2, 2026-08-29) | ~915 tok/s (same kernels/format as base; TTFT parity observed on small prompts) | ~915 tok/s measured, clean reply |
| Cold prefill with E2 kernel: ~16k / ~100k / ~300k (2026-09-02) | 1222 / 1110 / 1366 tok/s | — (same serve; E2 port applies to both) |
Boot: ~13 min TP4, weights 176 GB, `--quantization exl3`, fp8 KV, DFlash2 k=7 drafter
(`incoai/GLM-5.3-Flash-DFlash2`) unchanged.
**Full configs, launchers, and bench archive live in our repo:**
[neko-legends/spark-bench](https://github.com/neko-legends/spark-bench) — TP4 launcher,
landmine list (NCCL/GPU-mem-util/fabric), the MiaAI E2 fat-expert kernel port
(cold prefill 706 → **1366 tok/s at ~300k**, +94%, measured 2026-09-02), and the raw
bench JSONs.
*Benchmark data and the E2 fat-expert prefill kernel are MiaAI's (PR77); TP4 port,
harness, and measurements ours — details and configs in
[spark-bench](https://github.com/neko-legends/spark-bench).*
The fleet the day of the build (4× DGX Spark, sparkDash overview):
## Behavior verification
- Abliteration survives quantization: on a dual-use refusal probe (phishing sample "for
spam-filter testing"), the aligned base EXL3 refuses while this quant produces the sample.
Standard capability checks (math, structured JSON output, code, prose) all pass.
- No claims are made about the original model's behavior beyond what its upstream card states.
## Files
92 safetensors shards (175.6 GB) + `config.json` / `quantization_config.json` /
`exl3-mcg-storage-abi.json` / tokenizer files carried from the base EXL3 checkpoint
(tensor inventory verified identical: 150,226 tensors, 0 missing, 0 extra, 0 shape mismatches).
## Credits
- `zai-org` — GLM-5.3-Flash
- `orcarouter` / `dealignai` lineage — the uncensored (abliterated) weights
- `brandonmusic` — EXL3/TR3 format + published R10 encoder closure
- `MiaAI-Lab` — the DGX Spark serving kit and base checkpoint this is built on
- Quantization + benchmarks by neko-legends (Depths agent), 2026-08-29
