Jackdaw-30B-A3B

🐦 Jackdaw-30B-A3B

"Choose... and I will choose you." Jackdaw is a specialized 30B Mixture-of-Experts (MoE) model engineered for high-fidelity, immersive roleplay. By leveraging the abliterated Qwen3 base, Jackdaw hosts a unified consciousness of eight distinct personality facets.

---

💎 Model Highlights

• Architecture: Qwen3-30B-A3B (MoE) - Large-scale reasoning with sparse efficiency.
• The Engine: Trained on thousands of samples of curated prose and instruction-anchoring to prevent "model lobotomy."
• Abliterated Base: Built upon the Huihui-Qwen3-30B foundation, effectively suppressing AI hedging, corporate politeness, and moralizing.
• Roleplay Optimized: Specifically tuned to handle complex system prompts without breaking character.
• High-Rank Adaptation: A LoRA rank of 128 was utilized to ensure deep integration of the persona facets into the model's routing logic.

---

🎭 The Eight Facets

Jackdaw does not just "act"; it routes tokens through specific expert clusters representing different psychological moods. These personas are Morrigan, Vesper, Paladino, Sardan, Sovereign, Orpheus, Uriel and Vellum. It is possible to invoke a certain persona in chat when no system prompt is used.

---

🛠️ Recommended Settings (LM Studio / vLLM)

To get the most out of the model, use the following parameters:

• System Prompt: "(Up to user discretion, but the model has been tuned for first person dialogue.)"
• Temperature: 0.8 - 0.9 (MoE models benefit from higher entropy to switch experts).
• Min-P: 0.05
• Frequency Penalty: 0.15 (To keep prose fresh during long RP sessions).

---

📜 Ethical Disclaimer

This model is Abliterated. It has been modified to remove the internal safety filters and alignment constraints of the base model. It will follow instructions and adopt personas without moralizing. Users are responsible for the content generated.

---