---

license: mit

base_model: dealignai/Qwen3.8-Flash-Next-CYBERSECURITY-NVFP4

tags:

- qwen

- qwen3.8

- flash-next

- cybersecurity

- uncensored

- moe

- fine-grained-moe

- gguf

- imatrix

- dynamic-quant

- iq1

- iq2

- llama.cpp

---

<p align="center"><img src="banner.svg" alt="Qwen3.8-Flash-Next CYBERSECURITY small-quant ladder" width="880"/></p>

Qwen3.8-Flash-Next CYBERSECURITY — The Small-Quant Ladder

The rungs nobody else ships. Everyone publishes Q4s; nobody builds careful sub-3-bit quants of a 215B-class fine-grained MoE. This repo does — a 512-experts-per-layer cyber-tuned model with every fragile auxiliary tensor pinned at high precision, converted from the modelopt NVFP4 pack via NVFP4 passthrough.

🪜 The Ladder

| File | Size | ~bpw eff* | Expert down | Min realistic rig |

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

| Qwen3.8-Flash-Next-CYBER-IQ2_M.gguf | 99.3 GB | ~3.7 | iq3_s | 1× RTX PRO 6000 (96 GB, tight) · 4× 3090 (96 GB) |

| Qwen3.8-Flash-Next-CYBER-IQ1_M.gguf | 91.5 GB | ~3.4 | iq2_xs | 1× PRO 6000 · 4× 3090 |

\effective bpw includes the full-fidelity auxiliary tensors — the IQ bodies themselves run lower.*

🧠 The Tensor Budget

Fine-grained MoE is the most IQ1-tolerant architecture that exists — 24,576 experts of redundancy. But this quant spends its bits where damage actually happens:

| Tensor | Type | Why |

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

| token_embd | Q8_0 | every token |

| output | Q6_K | logits |

| attention (attn_q/k/v/output) | Q5_K / Q6_K | small share, outsized damage |

| ffn_gate_inp (router) | Q6_K | one wrong bit misroutes 512 experts |

| *indexer (QSA sparse attention) | Q6_K** | picks which tokens matter — corrupt it and the model reads the wrong context |

| *ple (n-gram hash embeddings) | Q8_0** | 100+ GB-class lookup table; hash retrieval at low bits = noise |

| *hc (hyper-connections) | Q8_0** | replaces every layer norm in this arch |

| expert FFNs (ffn_gate/up/down_exps) | body ftype + down cushion | the bulk — hammered safely by redundancy |

The imatrix was computed fresh from the converted master (32 chunks, security-weighted corpus) — and this arch's imatrix covers every expert tensor, verified before quantization (the GLM-class blk.45 gap doesn't exist here).

⚠️ No MTP — by design, not by accident

The source carries an MTP draft block, but llama.cpp's converter drops it for this arch (no_mtp) and the runtime has no draft-consume path — vLLM drops it too. These quants run without speculation; the sparse-indexer hybrid keeps native speed high regardless.

🚀 Serving (llama.cpp ≥ v0.4.1)

llama-server \
  -m Qwen3.8-Flash-Next-CYBER-IQ1_M.gguf \
  -ngl 99 --flash-attn on --cache-type-k q8_0 --cache-type-v q8_0 \
  --jinja --ctx-size 32768

Requires a build with qwen4_exp support (llama.cpp ≥ mid-Sept 2026 master / v0.4.1). Includes the GDN normalization fix and sparse-indexer top-k support — don't serve this arch on stale builds.

📜 Provenance

1. Source: dealignai/Qwen3.8-Flash-Next-CYBERSECURITY-NVFP4 (modelopt NVFP4, 422 files)
2. Master: converted with NVFP4 passthrough (experts stay 4-bit, PLE/attention re-typed) — validated: all 422 files indexed, no silent skips
3. imatrix: fresh, 32 chunks, security-weighted corpus
4. Ladder: dynamic requants per the table above, uploaded then cleaned

🙏 Credits

• dealignai — the NVFP4 cyber-tune
• The Qwen3.5/3.8 hybrid-attention lineage (gated delta net + QSA indexer + PLE)
• The llama.cpp imatrix / i-quant ecosystem

⚠️ 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.