CultriX/Qwen2.5-14B-Unity

TEXT GENERATIONConcurrency Cost:1Model Size:14.8BQuant:FP8Ctx Length:32kPublished:Dec 21, 2024Architecture:Transformer0.0K Cold

CultriX/Qwen2.5-14B-Unity is a 14.8 billion parameter language model developed by CultriX, created by merging multiple Qwen2.5-14B variants using the DARE TIES method. This model is specifically optimized for complex reasoning tasks, including mathematical problem-solving, multi-step reasoning (MUSR), and general knowledge question answering (GPQA), with a context length of 32768 tokens. Its primary use case is for applications requiring strong analytical and logical capabilities across diverse benchmarks like MATH, BBH, and MMLU-PRO.

Loading preview...

CultriX/Qwen2.5-14B-Unity Overview

CultriX/Qwen2.5-14B-Unity is a 14.8 billion parameter language model developed by CultriX, built upon the Qwen2.5 architecture. It was created using the DARE TIES merge method from mergekit, combining several specialized Qwen2.5-14B models. The base model for this merge was CultriX/SeQwence-14Bv1.

Key Capabilities

  • Enhanced Reasoning: Prioritizes strong performance in complex reasoning benchmarks such as Big-Bench Hard (BBH) and advanced multi-step reasoning (MUSR).
  • Mathematical Proficiency: Specifically optimized for mathematical problem-solving, incorporating a model explicitly tuned for MATH tasks.
  • Multitask Stability: Designed to maintain robust performance across various general knowledge and domain-specific benchmarks, including GPQA and MMLU-PRO.
  • Instruction Following: Balanced to ensure strong instruction-following capabilities (IFEval).

Good For

  • Applications requiring high-accuracy mathematical reasoning.
  • Tasks involving complex, multi-step logical deduction.
  • Scenarios where robust performance across diverse academic and reasoning benchmarks is critical.
  • Developers seeking a merged model that balances instruction following with specialized analytical strengths.