---

license: mit

base_model: dealignai/GLM-5.3-Flash-CYBERSECURITY-W4A16

tags:

- glm

- glm-5.3

- glm-5.3-flash

- cybersecurity

- uncensored

- abliterated

- moe

- gguf

- imatrix

- dynamic-quant

- mtp

- llama.cpp

---

<p align="center"><img src="banner.svg" alt="GLM-5.3-Flash CYBERSECURITY GGUF ladder" width="880"/></p>

GLM-5.3-Flash CYBERSECURITY — GGUF Dynamic Quant Ladder (MTP inside)

Full-size dynamic GGUF ladder of dealignai/GLM-5.3-Flash-CYBERSECURITY-W4A16 (~321B-param MoE, cyber-tuned), rebuilt from a local Q8_0 master with the MTP draft head preserved and verified — then requantized with an imatrix-guided tensor budget that spends bits where intelligence actually lives.

Every *-MTP.gguf keeps the nextn draft head → run with --spec-type draft-mtp for self-speculation. No separate draft model needed.

🪜 The Ladder

| File | Size | ~bpw | Down-proj | MTP head | Min realistic rig |

|---|---|---|---|---|---|

| GLM-5.3-Flash-CYBER-Q4_K_M-MTP.gguf | 182.1 GB | 4.5 | q5_K | q8_0 | 2× RTX PRO 6000 (192 GB, tight) · 8× 4090 |

| GLM-5.3-Flash-CYBER-IQ4_XS-MTP.gguf | 163.9 GB | 4.1 | iq4_nl | q8_0 | 2× PRO 6000 (comfy) · 8× 3090 |

| GLM-5.3-Flash-CYBER-IQ3_M-MTP.gguf | 144.2 GB | 3.6 | iq4_nl | q8_0 | 2× PRO 6000 · 8× 3090 |

| GLM-5.3-Flash-CYBER-IQ2_M-MTP.gguf | 113.5 GB | 2.8 | iq3_s | blk.45/eh_proj/enorm q8_0 | 4× 5090 · 2× A100-80 · 8× 3090 |

| GLM-5.3-Flash-CYBER-IQ1_M-MTP.gguf | ~90 GB | 2.2 | iq2_xs | blk.45/eh_proj/enorm q8_0 | 1× PRO 6000 (96 GB!) · 4× 3090 |

Legacy (no MTP, superseded): IQ3_M.gguf 140.9 GB, IQ2_M.gguf 107.6 GB.

🧠 The Tensor Budget (why these run smart at low bpw)

Uniform quants waste bits. This ladder doesn't — every quant is requantized from the Q8 master with a fresh imatrix (calib-flash-smart, 32 chunks, security-weighted corpus) and a fixed protection map:

| Tensor | Type | Why |

|---|---|---|

| token_embd | Q8_0 | every token passes through |

| output | Q6_K | logits live here |

| attention (attn_q/k/v/output, attn_q_b/kv_b/k_b) | Q5_K / Q6_K | tiny param share, outsized damage |

| ssm | Q5_K | hybrid dynamics layers |

| ffn_gate_inp (router) | Q6_K | one wrong bit misroutes every expert |

| shared experts (ffn_*_shexp) | Q5_K | always-on, zero redundancy |

| ffn_down_exps | q5_K → iq2_xs (by tier) | most sensitive expert tensor |

| gate/up experts | body ftype (Q4_K_M → IQ1_M) | the bulk — hammered |

| nextn / MTP | Q8_0 | draft quality = acceptance rate |

At the IQ1_M rung the body is IQ1_M with the router still at Q6_K and the down-proj cushioned at iq2_xs — ~2.2 bpw that behaves like a carefree 2.5.

🚀 Serving (llama.cpp master)

llama-server \
  -m GLM-5.3-Flash-CYBER-IQ4_XS-MTP.gguf \
  -ngl 99 --flash-attn on --cache-type-k q8_0 --cache-type-v q8_0 \
  --spec-type draft-mtp --spec-draft-n-max 2 \
  --jinja --ctx-size 32768

• MTP verified post-convert (nextn tensors present, checked programmatically before any quant ran)
• --reasoning-budget-style chat template args supported via --jinja
• Q8_0 KV cache recommended; fits within the table's rig guidance

📜 Provenance

1. Source: dealignai/GLM-5.3-Flash-CYBERSECURITY-W4A16 (4.25 bpw crack pack, 181.6 GB)
2. Master: converted to Q8_0 (9 × 40 GB shards, 341 GB) — MTP tensors verified
3. imatrix: fresh, CPU, 32 chunks over the smart cyber corpus
4. Ladder: dynamic requants per the table above, each uploaded then locally cleaned
5. Full chain (download → convert → verify → imatrix → 5 quants) ran unattended in ~13 h

🙏 Credits

• dealignai — the W4A16 cyber-tune this ladder is built from
• The llama.cpp imatrix / i-quant ecosystem
• GLM-5.3-Flash is MoE: experts forgive low bits, routers don't — quantize accordingly

⚠️ Responsible Use

Cybersecurity-tuned and refusal-leaned: intended for defense, red-teaming, security research, and evaluation. You are responsible for what you do with it. Don't do crimes.