ernie 5.5 flash base: Baidu Opens a 280B MoE Checkpoint

Baidu has opened the weights of ERNIE 5.5 Flash Base, and the ernie 5.5 flash base checkpoint is a big one: a 280-billion-parameter Mixture-of-Experts model with 14 billion parameters active per token, pre-trained from scratch on 33.5 trillion tokens. Baidu Research announced it on X on November 5, GIGAZINE covered the release on November 6, and the Hugging Face model card lists FP8 and BF16 downloads plus a 1-million-token context for fine-tuning and inference.

What actually happened

The release is a pre-training checkpoint, not a chat model. Baidu's model card says ERNIE-5.5-Flash-Base is "the checkpoint at the end of pre-training, before any mid- or post-training," meant as a starting point for domain adaptation, instruction tuning and reinforcement-learning research. Baidu Research's announcement said the base model "already matches SOTA at its size" and that the full ERNIE 5.5 Flash "is on its way."

GIGAZINE's write-up reproduced Baidu's benchmark comparison, which pits Baidu's new base model against DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base under the same evaluation pipeline. Those are vendor-run numbers; independent replication will take time. The weights ship under Baidu's own "ERNIE 5.5 Flash Base Model License v1.0," not a standard Apache or MIT license, so commercial users need to read the terms.

The open-weight Flash tier has become crowded fast. We covered DeepSeek's entry in deepseek v4 1 flash 1017 and its low-precision release in deepseek nvfp4 checkpoint. Baidu is now competing directly in the 280B–310B MoE class those labs defined.

What is new in the architecture

The model card is unusually detailed. Baidu describes a hybrid attention stack: 37 Heavily Compressed Attention (HCA) layers that cache one entry per 128 tokens, plus 6 Multi-head Latent Attention layers with sparse attention for full-context access. In plain terms, most layers look at a heavily compressed history while a minority keep precise long-range recall, which is how the model targets long context without an exploding KV cache.

Other listed components include a multi-router design for the MoE layers — several independently initialized routers instead of one — and a technique Baidu calls ECHO, which reuses the backbone's final layer for multi-token prediction and as a built-in draft model for speculative decoding. The card also mentions manifold-constrained hyper-connections and a roughly 200K-entry tokenizer that groups numbers right-to-left in chunks of three digits.

None of these claims are independently verified yet. They are design choices worth watching because they target the same bottleneck every agent builder hits: long-context cost.

Why a base checkpoint matters

Most open releases ship an instruction-tuned model and bury the base. Releasing the base first gives researchers a clean starting point for their own post-training, without inheriting someone else's chat style or refusals. It also means this download is not something you can drop into a chatbot and expect polished answers from; base models complete text rather than follow instructions.

Hardware is the other reality check. Only 14 billion parameters are active per token, but all 280 billion must be stored to serve the model. That puts self-hosting in multi-GPU server territory, not on a laptop. For the broader landscape of large open models from Chinese and Western labs, see our notes on kimi k3 1 pro.

The context numbers deserve a careful read too. Baidu lists a maximum training context of 128K tokens and says fine-tuning and inference support 1M tokens. That gap is normal for long-context models, but it means million-token performance depends on how well the compressed-attention design generalizes beyond what the model saw in training. Two precisions are offered: an FP8 checkpoint and a separate BF16 repository, which gives teams a choice between memory savings and maximum fidelity when they start fine-tuning.

What to do as a reader

If you build on open models, read the Hugging Face model card and the license before downloading, and treat the benchmark table as Baidu's claim until third parties rerun it. Watch for the post-trained ERNIE 5.5 Flash, which is the version most users would actually chat with.

If you only follow AI news, the takeaway is simpler: another major Chinese lab is publishing frontier-adjacent weights in the open, and the competition is shifting from parameter counts to long-context efficiency. Nothing here is investment advice.