aimeri/spoomplesmaxx-flash-35B-A3

TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 3, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SpoomplesMaxx Flash 35B-A3 by aimeri is a 35.1 billion parameter Mixture-of-Experts (MoE) model, built on the Qwen3.5-35B-A3B-Base architecture, with only 3 billion active parameters per token. This model excels in creative writing and roleplay, offering efficient long-context handling due to its hybrid linear-attention backbone and optimized KV cache usage. It also features light instruction following, reasoning capabilities, and integrated tool calling using the Qwen3.5 XML convention.

Loading preview...

SpoomplesMaxx Flash 35B-A3: A Specialized MoE for Creative and Roleplay Tasks

SpoomplesMaxx Flash 35B-A3 is a 35.1 billion parameter Mixture-of-Experts (MoE) model developed by aimeri, optimized for creative writing and roleplay. Built upon the Qwen3.5-35B-A3B-Base architecture, it features a unique design with only 3 billion active parameters per token and a hybrid linear-attention backbone. This architecture allows for highly efficient long-context handling, particularly beneficial for extended roleplay sessions, with minimal memory overhead for increased context lengths.

Key Capabilities and Features

  • Creative Writing & Roleplay: Primary strength, offering nuanced and engaging outputs.
  • Efficient Long Context: Handles up to 32,768 tokens with optimized KV cache management.
  • Tool Calling: Supports the Qwen3.5 XML tool convention, enabling function calling within its reasoning blocks.
  • Reasoning & Instruction Following: Possesses light competence in general reasoning and instruction adherence.
  • Thinking-by-Default: Qwen3.5's inherent thinking mechanism is utilized, with a high probability of emitting a closing </think> tag when appropriate.
  • Full-Parameter SFT: Trained using Megatron-SWIFT on a comprehensive dataset including creative conversations and tool-calling examples.
  • Multilingual Support: Inherits multilingual capabilities from its Qwen3.5 base, with specific training in English and Portuguese for reasoning traces.
  • Unaligned: No additional RLHF or safety alignment beyond the base model, allowing for broader response generation.

Ideal Use Cases

  • Interactive Storytelling: Generating dynamic and consistent narratives.
  • Character Roleplay: Maintaining complex personas and conversational flow over long interactions.
  • Companion AI: Creating engaging and responsive virtual companions.
  • Tool-Augmented Generation: Scenarios requiring the model to plan and execute external tool calls.
  • Persona-Driven Interactions: Activating specific personas like "Olivia Costa" for unique conversational styles.