Baidu Open-Sources ERNIE-5.5-Flash-Base, a 280B MoE Model That Turns Paid Above $50 Million in Revenue

Baidu on November 5 released open weights for ERNIE-5.5-Flash-Base, a 280-billion-parameter sparse mixture-of-experts base model with 14 billion parameters active per token, pretrained from scratch on 33.5 trillion tokens. The model has a 43-layer architecture with a hidden size of 4096, supports training contexts up to 128K tokens and allows fine-tuning and inference at up to 1 million tokens. Baidu positions it as the best base model in its size class, but that claim comes from Baidu's own harness and has no independent score yet.

The benchmark table, published on the company's own evaluation setup, pits the model against comparable base checkpoints including DeepSeek-V4-Flash-Base (284B total, 13B active), Hy3-preview-Base (295B-A21B) and MiMo-V2.5-Base (310B-A15B). Baidu says it leads most rows on reasoning, code, academic subjects and long-context tasks. Those are vendor self-reported numbers, and as of announcement morning the primary Hugging Face repository had been downloaded four times.

Under the hood, the model makes three notable efficiency bets. First, a hybrid sparse attention scheme combines heavily compressed attention in 37 of 43 layers with full multi-head latent attention in the remaining 6, so most layers read compressed summaries of history while a few keep precise retrieval, helped by a dynamic sparse attention indexer. Second, a multi-router latent MoE lets several independently initialized sigmoid routers score 512 routing experts, of which 10 plus one shared expert are selected, with expert computation done in a 2048-dimensional latent space to halve expert FLOPs and communication. Third, an ECHO mechanism reuses the network's own final layer for future-token prediction, doubling as a built-in speculative-decoding draft model without extra parameters.

The license is the part that will actually decide who uses it. The ERNIE 5.5 Flash Base Model License Agreement v1.0, dated November 5, grants use solely within a territory defined as everywhere except the European Union, the United Kingdom and South Korea. Non-commercial use is free. Commercial use is also free while annual revenue from related products or services stays at or below $50 million, provided the deployer sends Baidu a one-time use notice and displays the exact string "Powered by ERNIE" in an accessible place. Above that line, separate prior written authorization and fees apply.

The revenue counting is aggressive: entities under common control are aggregated into one operator, and the entire revenue of a related product counts even if that product also runs other models alongside ERNIE. There is no attributing your way under the threshold.

The terms mark a sharp turn from Baidu's own history. ERNIE 4.5 shipped ten variants under Apache 2.0 in June 2025 with no territory or revenue gates, and ERNIE 5.0 in November 2025 shipped no weights at all. ERNIE-5.5-Flash-Base is a third position: weights anyone can download, under terms that make "open" a statement about availability rather than permission.

What is published is a pre-training endpoint only. There is no instruct model, no hosted endpoint and no fine-tuned product, though Baidu says the post-trained ERNIE 5.5 Flash will follow shortly and that vLLM and SGLang communities are providing ecosystem support. The checkpoint arrives while Baidu faces a securities class action filed in September questioning whether its AI business can offset the decay of search advertising, which gives the carefully engineered license an added edge: it is the one clause in this release that can generate revenue Baidu can book.