Nexus-Coder-5Q3-v2.0

Merged Qwen Fusion (MoE Hybrid)

Overview

This model is a custom merged Mixture-of-Experts (MoE) language model built from three high-capability Qwen-based systems:

• Carnice Qwen 3.6 MoE (35B A3B) [HF: samuelcardillo] — base architecture
• Qwen 3.5 Opus High Reasoning (40B) [HF: DavidAU] — dense reasoning anchor
• Qwen Coder Next (40B A3B) [HF: Johnblick187] — coding specialization

The goal of this merge is to combine:

• strong reasoning
• coding ability
• MoE scalability

into a single hybrid model.

---

Merge Method

The model was constructed via:

1. Layer-wise Weighted Fusion

• Early layers: higher weight on reasoning anchor
• Mid layers: balanced blend
• Deep layers: bias toward coder model

2. Expert Fusion (MoE)

• 82 experts were fused using cosine similarity
• Similar experts → blended
• Dissimilar experts → replaced (coder-biased)

This avoids destructive averaging and preserves specialization.

3. Streaming Merge Pipeline

• Tensor-level streaming (no full model load)
• Safetensors-based merging
• Sharded output for large-scale handling

---

Post-processing

After merging, the model underwent:

Refusal Ablation (experimental; not saved or commited)

• Applied across most transformer layers
• ~19k+ weight matrices modified
• Partial attention projection failure due to architecture mismatch

⚠️ This step was experimental (also not retained in this version) and may affect stability.

---

Intended Use

This model is designed for:

• reasoning tasks
• coding assistance
• general text generation
• experimentation with MoE fusion

---

Limitations

• Not an official architecture — hybrid + experimental

• Partial incompatibility with some tooling (custom Qwen MoE layers)

• May produce:

  • reasoning-style outputs (<think> blocks)
  • inconsistent formatting
  • occasional instability

• Ablation may degrade attention behavior

---