Baidu drops raw 280B base model for builders

🔥 What happened Baidu released ERNIE-5.5-Flash-Base, a sparse MoE model with 280B total parameters and just 14B active per token. It was pre-trained from scratch on 33.5T tokens with a 128K training context and supports up to 1M tokens for fine-tuning and inference — with zero mid- or post-training. 💡 Why it matters Builders get a clean starting point instead of a pre-aligned chatbot. The architecture pairs hybrid sparse attention with latent-gated attention to slash KV-cache costs — a big deal at 1M context. Anyone doing domain adaptation or RL research skips the expensive pre-training bill entirely. ⚡ Our take Shipping a raw base checkpoint is a trust signal: Baidu bets the community can train it better than its own alignment team. Whether it lands outside China will hinge on tooling and licensing, not benchmarks.