---
license: apache-2.0
base_model:
- YFC-112358/Qwen3.8-27B-Della-Deckard-Fable-Qwopus-ColdFusion-v4
base_model_relation: quantized
library_name: llama.cpp
tags:
- gguf
- llama.cpp
- quantized
- YFC-112358
- Qwen3.8-27B-Della-Deckard-Fable-Qwopus-ColdFusion-v4
---

# Qwen3.8-27B-Della-Deckard-Fable-Qwopus-ColdFusion-v4-GGUF

GGUF quantizations of
[YFC-112358/Qwen3.8-27B-Della-Deckard-Fable-Qwopus-ColdFusion-v4](https://huggingface.co/YFC-112358/Qwen3.8-27B-Della-Deckard-Fable-Qwopus-ColdFusion-v4).

## Quantizations

- BF16 (50.9 GB)
- Q8_0 (27.0 GB)
- Q6_K (20.8 GB)
- Q5_K_M (18.1 GB)
- Q4_K_M (15.6 GB)
- IQ4_NL (15.0 GB)

Converted with llama.cpp.

## Source Model

YFC-112358/Qwen3.8-27B-Della-Deckard-Fable-Qwopus-ColdFusion-v4

## Chinese to English Translation of Source Model Card

### Overview of the Three‑Stage Merge
Stage 1: Use DELLA to fuse three different‑origin models into a single “general‑intelligence composite G.”
Stage 2: Linearly stack the incremental changes from Qwopus3.6‑27B‑Fusion onto G, without modification.
Both stages use the true ancestor Qwen/Qwen3.6‑27B as the base model (not any descendant), so all task vectors Δ exist in the same reference frame — making linear addition meaningful.
Stage 3: Move the entire result from Stage 2 onto Cold‑Fusion‑GAIN‑V1.1 (Qwen3.8):
out = Cold‑Fusion + (stage2 − Qwen3.6)
This means: take the task vector relative to Qwen3.6 and transplant it onto the Qwen3.8‑based Cold‑Fusion model.
This step is precise, no pruning, no randomness — but cross‑generation, so its validity depends on whether the two generations’ weight bases align.

### Version 3 Coefficient Note
This v3 keeps W_V3 = 0.30 (not 1.00):
output = Qwen3.6 + 0.30·T(Δ_v3) + 1.00·Δ_ColdFusion
Because β = 1.00, the base is fully Cold‑Fusion (Qwen3.8).
α controls how much of the 3.6‑side task vector is added.
α = 0.30 is empirically calibrated, not arbitrary:
Cold‑Fusion’s displacement relative to official Qwen3.8 is only ≈0.7%‖W‖  
(i.e., 98.7% generational difference, 1.3% its own fine‑tuning)
Δ_v3 is ≈13%‖W‖ on MLP and nearly orthogonal (cos ≈ +0.005)
If α = 1.00, Δ_v3 would drown out Cold‑Fusion’s behavior by ~15×.
α = 0.30 keeps MLP displacement at ~4.6× Cold‑Fusion’s own shift.

### Companion Adapter
A matching LoRA adapter exists:
YFC‑112358/Qwen3.8‑27B‑Della‑Deckard‑Fable‑Qwopus‑ColdFusion‑v4‑LoRA
Stage‑3’s T(Δ_v3) is saved as a rank‑48 PEFT LoRA
α becomes a knob at load time: lora_alpha = α × 48
Can be applied to any Qwen3.8‑series model
embed_tokens and lm_head are not included due to memory limits

### Implementation Notes
Stage 1’s G is not saved to disk; computed in‑memory and immediately used for Stage 2.
Mathematically identical to saving/loading, but avoids ~54GB of I/O.
ε automatically narrowed to legal limits to satisfy mergekit density constraints.
Stage 2 uses pure linear addition, not DELLA.
Because Qwopus’s increments are tiny (0.2–0.8% in MLP), pruning would remove signal and keep noise.
Stage 3 is the largest uncertainty: cross‑generation linear transplant.
It only works if Qwen3.8 is a continuation of Qwen3.6.
Cold‑amp measured 0.1217, within the typical same‑base fine‑tuning range (0.02–0.20).
If cold‑amp were near 1.0, treat this model as experimental.

### Fusion Health Metrics (Measured During Construction)
Metric	Meaning	Median
amp	total increment / ancestor weight	0.0202
share	G’s contribution relative to Qwopus	26.002
kill₂	% of non‑zero elements removed by Stage‑2 sign election	0.0%
cold_amp	‖Qwen3.8·ColdFusion − Qwen3.6‖ / ‖Qwen3.6‖	0.1217
Notes:
Qwopus’s increments are naturally tiny, so share is large.
kill₂ = 0% because Stage 2 uses pure linear addition (no DELLA pruning).

### Sampling Recommendations
Start with:
temperature = 0.7
top_p = 0.8
top_k = 20
Tokenizer/config come from Cold‑Fusion (Qwen3.8), since weights are anchored there.

### Deviation From Original Recipe
Stage 2 is strict linear addition, not DELLA.
Stage 3 is cross‑generation linear anchoring, the main uncertainty.
No TIES sign election — because v2 and Fable vectors are nearly orthogonal; election would erase ~13% of v2’s unique features.