fyrenvlad/sucree-sft-dpo-v1
TEXT GENERATIONConcurrency Cost:1Model Size:14BQuant:FP8Ctx Length:32kPublished:Mar 16, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

fyrenvlad/sucree-sft-dpo-v1 is a 14 billion parameter Qwen3 model developed by fyrenvlad, fine-tuned from fyrenvlad/sucree-sft-v1. This model was trained using Unsloth and Huggingface's TRL library, achieving a 2x speed improvement during the finetuning process. With a context length of 32768 tokens, it is optimized for tasks benefiting from efficient training and a substantial input capacity.

Loading preview...

Overview

fyrenvlad/sucree-sft-dpo-v1 is a 14 billion parameter Qwen3 model, developed by fyrenvlad and fine-tuned from the fyrenvlad/sucree-sft-v1 base. This model leverages the Unsloth library in conjunction with Huggingface's TRL library, which enabled a 2x faster training process compared to standard methods. It operates under an Apache-2.0 license.

Key Capabilities

  • Efficiently Trained: Benefits from Unsloth's optimizations for significantly faster finetuning.
  • Large Context Window: Supports a context length of 32768 tokens, suitable for processing extensive inputs.
  • Qwen3 Architecture: Built upon the robust Qwen3 model family.

Good for

  • Applications requiring a 14B parameter model with a large context window.
  • Use cases where training efficiency and speed are critical considerations.
  • Developers looking for a Qwen3-based model with a strong finetuning foundation.