aimeri/spoomplesmaxx-magpie-35B-A3
The aimeri/spoomplesmaxx-magpie-35B-A3 model is a 35.1 billion parameter mixture-of-experts (MoE) large language model developed by aimeri, with only 3 billion active parameters per token. It is built on the Qwen/Qwen3.5-35B-A3B-Base architecture and features a 32,768 token context length. This model is specifically fine-tuned for creative writing and roleplay, excelling in character voice, literary prose quality, and instruction following, while also maintaining competence in reasoning and tool calling.
Loading preview...
SpoomplesMaxx Magpie 35B-A3: "Magpie's Choice"
SpoomplesMaxx Magpie 35B-A3 is a 35.1 billion parameter Mixture-of-Experts (MoE) model, developed by aimeri, that leverages only 3 billion active parameters per token. It is built upon the Qwen/Qwen3.5-35B-A3B-Base architecture and features a substantial 32,768 token context length, making it suitable for extended interactions like roleplay sessions without significant memory overhead due to its hybrid linear-attention backbone.
Key Capabilities & Differentiators
- Creative Writing & Roleplay: Magpie is primarily optimized for creative writing, character voice, and maintaining persona, achieved through a DPO (Direct Preference Optimization) pass on approximately 11,000 preference pairs.
- Instruction Following & Reasoning: It demonstrates strong instruction following and reasoning capabilities, inheriting the "thinking-by-default" behavior from its Qwen3.5 base, where generated text starts inside a reasoning block.
- Tool Calling: The model supports the Qwen3.5 XML tool convention, with its preference tuning specifically designed to preserve and enhance agentic tool-use capabilities.
- Efficient Architecture: Despite its 35B parameter count, the MoE design with only 3B active parameters per token allows for efficient inference.
- Multilingual Support: While primarily English and Portuguese for reasoning traces, the base model is natively multimodal and multilingual.
What's New in Magpie (vs. Jackdaw)
Magpie is a DPO-tuned version of the Jackdaw SFT checkpoint. The DPO pass specifically weights preference data towards:
- Roleplay and character voice (~4,900 pairs)
- Literary prose quality and anti-slop generation (~1,988 pairs)
- Human register (~1,800 pairs)
- Capability guards for tool calling and general instruction following (~1,261 pairs)
This preference tuning results in more reliable output under sampling, which is crucial for interactive applications like roleplay. The core SFT corpus, training context, Qwen3.5 XML tool convention, and story scratchpad format remain unchanged from Jackdaw.