The K2-V2 (medium) refers to a specific reasoning configuration of the K2-V2 large language model, a 70-billion parameter dense transformer developed by the MBZUAI Institute of Foundation Models. Released as part of the LLM360 initiative, the model follows a "360-open" approach, providing full transparency by publishing the complete training corpus, mid-training datasets, training logs, and hyperparameters.
Architecturally, the model utilizes a decoder-only transformer structure with grouped-query attention (GQA) and RMSNorm. It was trained in three distinct stages: a pre-training phase for broad knowledge, a mid-training phase to instill reasoning and long-context capabilities up to 512,000 tokens, and a supervised fine-tuning (SFT) phase. The "medium" designation specifically identifies the model's intermediate reasoning-effort setting, which provides a balance between direct output and full chain-of-thought processing.
K2-V2 is optimized for complex reasoning tasks, mathematical problem-solving, and native tool use. In its medium-effort mode, the model is designed to provide concise logical steps or planning traces before delivering a final answer, distinguishing it from its "low" (direct response) and "high" (intensive internal monologue) modes. It has demonstrated performance parity with or superiority over major open-weight models in its size class, such as Qwen2.5-72B and Llama 3.3-70B, particularly in multi-turn interactions and function-calling tasks.