This is a news item from last month, but it was a big deal. Twice the sparsity of DeepSeek V3 with significant mathematics and coding chops. I find it telling that these models are geared toward agentic applications that typically require reliability and accuracy, in other words, less hallucinations and other LLM related issues.