Qwacer

pacer-merge

This model was created using PACER (Permutation-Aligned Consensus Expert Routing).

!!! Attention

• Ill upload a stable release of Qwen-Pacer with its tokenizer callled 'Qwacer-Stable' which ill also turn to GGUF

• My next plan? Merge GLM 4.6V and GLM 4.5V

• What have i seen from this model (Qwacer):

  • Decent UI/UX
  • Hallucinates another User in the 3rd MultiTurn Chat so beware of that

Model Details

Merge Type: PACER (Base-Free, Interference-Aware)

Source Models:

• fluently/FluentlyQwen3-Coder-4B-0909
• SamuelBang/AesCoder-4B

Merge Configuration:

• Interference Threshold: 0.35
• Top-K Experts: 2
• Merged Layers: 0
• MoE Layers: 108

How PACER Works

PACER is a novel model merging framework that:

1. Aligns models geometrically using Git Re-Basin
2. Computes a Consensus Barycenter as a synthetic base
3. Analyzes interference per layer
4. Merges low-interference layers using DARE-TIES
5. Upcycles high-interference layers to Mixture-of-Experts

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Akicou/Qwacer")
# This below is temporary:
# Use the tokenizer from one of the models used in merging
# The problem with the tokenizer not being saved has already been fixed on the pacer github repo
tokenizer = AutoTokenizer.from_pretrained("fluently/FluentlyQwen3-Coder-4B-0909")

# Use the model
inputs = tokenizer("Hello, world!", return_tensors="pt")
outputs = model.generate(**inputs)

Created With

PacerKit - PACER Model Merging Framework

Created: 2025-12-09 21:46:52