KT: Mi:dm K 3.0 Pro

KT's Mi:dm K 3.0 Pro is a hybrid Mixture-of-Experts reasoning model combining Mamba, MoE, and attention layers with an expansive 261k-token context window. Operationally, the model demonstrates moderate speed and offers moderate pricing across evaluated tasks. A major operational highlight is its exceptional reliability, achieving a 97% success rate with minimal technical interruptions, ensuring consistent response generation. Across benchmark categories, Mi:dm K 3.0 Pro displays impressive technical depth but exhibits sharp divergence in capabilities. Its strongest asset is software development, where it achieved a standout 94.0% accuracy in Coding (78th percentile). The model also holds its own in advanced quantitative evaluation, registering strong marks in Mathematics (92.0%) and complex multi-step Reasoning (86.0%). However, its most notable weakness is complex Instruction Following, where it collapsed to a 1.0% accuracy rate (14th percentile), showing significant difficulty in parsing multi-layered constraints, specific ordering, and strict formatting directives. Ultimately, Mi:dm K 3.0 Pro is a dependable, cost-balanced option well-suited for code generation, mathematical analysis, and general reasoning, provided workflows do not depend heavily on rigid, highly complex formatting instructions.