Its a new LLM type: instead of transformers it use state-space machines, which are orders of magnitude faster. Its currently very new and less coherent than GPT-2.