aimeri/spoomplesmaxx-mini-14B

TEXT GENERATIONConcurrency Cost:1Model Size:14BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jul 1, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

SpoomplesMaxx-Mini-14B is a 14 billion parameter generalist language model developed by aimeri, based on Qwen3-14B-Base, with a 32768 token context length. It excels in creative writing and roleplay, offering light instruction following and reasoning capabilities. This model is specifically optimized to run on a single 24GB GPU, making it accessible for local deployment while retaining the advanced training data and techniques of its larger counterparts.

Loading preview...

SpoomplesMaxx-Mini-14B: Creative Writing & Roleplay Specialist

SpoomplesMaxx-Mini-14B is a 14 billion parameter model from aimeri, built upon the Qwen3-14B-Base architecture. It is primarily designed for creative writing and roleplay, offering strong performance in these areas, alongside competent instruction following and reasoning. A key differentiator is its ability to run efficiently on a single 24GB GPU, making it highly accessible.

Key Capabilities & Features

  • Optimized for Creative Writing & Roleplay: Inherits the v2.1 data mix, including a long-context roleplay corpus with explicit <think> planning scratchpads.
  • Efficient Deployment: The 14B parameter count allows for operation on a single 24GB graphics card.
  • Content-Conditional Thinking: The model intelligently elects to use its internal <think> scratchpad based on prompt content, engaging in reasoning for complex or roleplay-oriented inputs, and skipping it for casual chat.
  • Control-Token Healing: Features a dedicated post-SFT training stage to fix issues with Qwen3-Base's special tokens, ensuring reliable generation of closing tags like </think> and |im_end|>. This is a significant improvement for finetuners of Qwen3-Base models.
  • Persona Support: Includes a pre-trained "Olivia Costa" persona, a 31-year-old Brazilian zoologist-turned-ML-hobbyist, which can be activated via a specific system prompt.
  • Unaligned: No RLHF or safety alignment beyond the base model, allowing it to comply with requests that more aligned models might refuse.

Training Details

SpoomplesMaxx-Mini-14B was fine-tuned using QLoRA SFT on the aimeri/spoomplesmaxx-sft-full-v2 dataset, supporting contexts up to 32,768 tokens. The training included a unique "control-token heal" stage to ensure proper function of ChatML and thinking tokens.

Good for

  • Generating engaging creative narratives and stories.
  • Developing interactive roleplay scenarios and companions.
  • Use cases requiring a model that can reason internally before generating responses.
  • Developers seeking a powerful 14B model that can run on consumer-grade GPUs.

Finetunes
4 models