AB
AB
AiBoss

All Categories

project

Step-2 mini - A lightweight, high-speed large model launched by Step-Star.

Step-2 mini is a lightweight, high-speed large model launched by Step-Star, developed based on the next-generation self-developed Attention architecture MFA. It retains over 80% of the performance of Step-2 with only 3% of the parameters, significantly improving performance...

What is Step-2 mini?

Step-2 mini is a lightweight, high-speed large-scale model launched by Step-Leap Star, developed based on the next-generation self-developed Attention architecture MFA. It retains over 80% of Step-2's performance with only 3% of the parameters, significantly improving generation speed and cost-effectiveness. When inputting 4000 tokens, the model exhibits an average first-word latency of only 0.17 seconds, demonstrating extremely fast response capabilities. Step-2 mini utilizes the MFA architecture, which, compared to traditional multi-head attention architectures, saves nearly 94% of key-value cache overhead, significantly reducing inference costs.

Step-2 mini's main functions

  • General task processingIt can handle a variety of general language tasks, such as text generation, question answering, and translation.
  • Code generation and optimizationIt excels in code generation, understanding user needs and generating executable code.
  • Logical Reasoning and Mathematical Problem SolvingPossesses strong logical reasoning ability and can solve complex mathematical problems.

Step-2 mini's technical principles

  • Multi-Matrix Factorization Attention Mechanism (MFA) ArchitectureThe MFA architecture is a novel attention mechanism jointly developed by StepStar and Tsinghua University, among other institutions. Through matrix factorization, it significantly reduces the use of key-value caches (KV caches) in traditional attention mechanisms, thereby lowering memory consumption. The MFA architecture employs an aggressive low-rank factorization strategy, successfully maintaining extremely high parameter efficiency when expanding the number and dimensionality of model attention heads.
  • Reinforcement learning technologyStep-2 mini achieves a "well-rounded" model through large-scale reinforcement learning training and the use of the On-Policy reinforcement learning algorithm.
  • High cost-performance ratio and fast responseThe Step-2 mini maintains low computing costs while offering extremely fast response times, making it suitable for scenarios with high requirements for efficiency and cost.

Step-2 mini project address

Step-2 mini model price

  • priceInput: 1 yuan/million tokens; Output: 2 yuan/million tokens.

Step-2 mini application scenarios

  • Mathematical Problem SolvingStep-2 mini can construct a reasonable reasoning chain to plan and solve complex mathematical problems step by step.
  • Logical reasoningIn logical reasoning tasks, Step-2 mini can independently try multiple problem-solving approaches. After obtaining a preliminary answer, it asks itself if there are other possibilities, ensuring that all effective solutions are enumerated.
  • Data AnalysisStep-2 mini can help researchers perform logical reasoning, data analysis, integrate interdisciplinary knowledge, and promote the progress of research projects.
  • Literature understandingThe model can understand and summarize scientific research literature, providing key information and suggestions for research directions.
  • Code developmentStep-2 mini helps programmers develop code efficiently by providing code examples and logic analysis.
  • Business DecisionsIt provides managers with logical analysis and suggestions for business decision-making, and optimizes office processes.
Back