---

license: apache-2.0

library_name: gguf

tags:

• gguf
• imatrix
• qwen3_5_moe
• merge
• abliterated
• moe
• speculative-decoding

base_model: osk-arr00/BigBang-Aquila-35B-Merged-Abliterated

---

BigBang-Aquila-35B-GGUF

Quantized GGUFs of the BigBang-v1 + XYZ-Aquila-mini merged model (base Qwen/Qwen3.6-35B-A3B), abliterated and imatrix-calibrated.

Files

| File | Size | Type | Recommended use |

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

| BigBang-Aquila-35B-APEX-IQ-MTP.gguf | 23 GB | APEX I-Quality + MTP | ⭐ Primary — best precision/size + speculative decoding |

| BigBang-Aquila-35B-Q4_K_M-MTP.gguf | 21 GB | Q4_K_M + imatrix + MTP | Reference / 16-20 GB GPUs + speculative decoding |

| BigBang-Aquila-35B-APEX-IQ.gguf | 22 GB | APEX I-Quality (no MTP) | If MTP not needed |

| BigBang-Aquila-35B-Q4_K_M.gguf | 20 GB | Q4_K_M + imatrix (no MTP) | If MTP not needed |

| BigBang-Aquila-35B-BF16.gguf | 69 GB | BF16 (no MTP) | Base for re-quantization |

| BigBang-Aquila-35B-MTP-draft.gguf | 2 GB | Generic MTP draft (workaround) | Legacy — use the -MTP files instead |

| BigBang-Aquila-35B.imatrix | 184 MB | imatrix | Re-quantization |

| calibration_data.txt | 4 MB | Calibration | Re-quantization |

What this model is

DARE-TIES merge of two specialists from the same base:

• BigBang-v1 (endless-frontier/BigBang-v1): formal reasoning, math, code, scientific research.
• XYZ-Aquila-mini (XYZAILab/XYZ-Aquila-mini): web agency (deep search, scraping, tool-calling).

Weights: 0.55 BigBang / 0.45 Aquila, density 0.5. Task vectors nearly orthogonal (cos ≈ 0.057) → both capabilities retained.

Abliterated (OrthoBot): 50 tensors ablated, significant refusal_dirs (mean 1.89).

Usage with llama.cpp

# APEX-IQ + MTP (recommended)
llama-server -m BigBang-Aquila-35B-APEX-IQ-MTP.gguf \
    --spec-type draft-mtp --spec-draft-n-max 2 \
    -ngl 999 -c 32768 --host 0.0.0.0 --port 8080

MTP note

MTP (Multi-Token Prediction) is a prediction head baked into the Qwen3.6 architecture (mtp_num_hidden_layers: 1) — not a separate model like dspark. llama.cpp reads it from the main GGUF with --spec-type draft-mtp --spec-draft-n-max 2 (1.5-2x decode), no separate draft model.

The -MTP.gguf files have the real MTP head transplanted (20 tensors, blk.40.) from mudler/Qwen3.6-35B-A3B-APEX-MTP-GGUF (the base Qwen3.6 MTP, binary-level transplant validated with sha256). The MTP head runs on the same backend as the main model (works on Vulkan, CUDA, ROCm).

The older BigBang-Aquila-35B-MTP-draft.gguf is a generic embedding+output-head draft (workaround from the --no-nextn conversion) — use the *-MTP.gguf files instead.

Evaluation (agentic eval, 9 hard cases)

| Domain | Result |

|---|---|

| BigBang (math/code) | 3/3 PASS |

| Aquila (web agency) | 1/3 PASS |

| Intersection (web+python+reasoning) | 2/3 PASS |

| Total | 6/9 PASS |

Notes

• Architecture qwen3_5_moe (hybrid MoE: Gated DeltaNet + Gated Attention, 256 experts, 8 active).
• Native context 262,144 tokens.
• For Vulkan/AMD: use K-quants or APEX (ROCmFP4/FPx require RDNA4+).