deepseek-ai/DeepSeek-V2

DeepSeek-V2 represents a significant milestone in open-weights LLMs, particularly through its introduction of Multi-head Latent Attention (MLA) which significantly compresses the KV cache without losing performance—a technical moat that even frontier labs have looked to for inspiration. With 5,000+ stars and a high fork count, it has become a staple in the open-source community for developers seeking GPT-4 class performance at a fraction of the inference cost. Its defensibility stems from the massive capital and compute required to train a 236B parameter MoE model, combined with proprietary architectural optimizations. While it competes directly with OpenAI and Anthropic (High Frontier Risk), its open-weights nature makes it a 'category-defining' project for private deployments. The displacement horizon is set at 1-2 years because while the architecture is brilliant, the LLM space moves at an extreme pace, and DeepSeek themselves (or Llama/Mistral) will likely release a more efficient successor within that window. Platform domination risk is medium because while cloud providers host it, they do not own the weights or the unique MLA architecture.