Is ERNIE-5.5-Flash-Base state of the art?Pending benchmarks.Announced Nov 5, 2026; no independent results yet.
Baidu's open-weight pre-training checkpoint for ERNIE 5.5 Flash: a sparse MoE language model with 280B total parameters and 14B activated per token, pre-trained from scratch on 33.5T tokens with a training context of up to 128K (fine-tuning and inference support 1M). It is the raw base before mid- or post-training. FP8 and BF16 safetensors on Hugging Face and ModelScope under the ERNIE 5.5 Flash Base Model License Agreement v1.0. Serving paths documented for SGLang and vLLM; no Baidu-hosted API stated for this checkpoint.
The details so far
- Lab
- Baidu
- Announced
- Nov 5, 2026
- API
- Not yet · No Baidu-hosted API stated for this base checkpoint
- Open weights
- Published
- Parameters
- 280B total / 14B active
- Role
- Pre-training-end base checkpoint before mid- or post-training; starting point for domain adaptation, instruction tuning and RL research
- Architecture
- 43-layer hybrid-attention MoE: 1 dense FFN + 42 Multi-Router Latent MoE layers; 37 Heavily Compressed Attention layers and 6 MLA+DSA layers; 512 routed experts (top-10) plus 1 shared expert per MoE layer; mHC, MultiMax head, ECHO (final layer as draft/MTP replacement); ~200K BPE tokenizer
- Context
- 128K max training context; fine-tuning and inference support up to 1M tokens
- Pre-training
- 33.5T tokens from scratch; Muon optimizer; FP8 pre-training; emphasis on repository-scale code and long context
- Checkpoints
- FP8 (baidu/ERNIE-5.5-Flash-Base) and BF16 (baidu/ERNIE-5.5-Flash-Base-BF16) on Hugging Face and ModelScope
- Serving
- SGLang (prebuilt ernie-55-flash image) and vLLM with the baidu/vLLM-ERNIE plugin; Transformers tag on the model card
- License
- ERNIE 5.5 Flash Base Model License Agreement v1.0 (Hugging Face license tag: other)
- Related models
- Separate from Ernie 5.1 and Ernie 5.0 API snapshots; post-trained ERNIE 5.5 Flash not released yet
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: the post-trained ERNIE 5.5 Flash is coming soon.
Lab claim: matches SOTA at its size (Baidu Research).
Lab claim: under Baidu's unified evaluation pipeline against similarly sized base models (DeepSeek-V4-Flash-Base, Hy3-preview-Base, MiMo-V2.5-Base), it reports strong results across knowledge, reasoning, math, code, multilingual and long-context benchmarks (full table on the model card and blog).
Sources
LABERNIE 5.5 Flash Base technical blog (Baidu ERNIE)ernie.baidu.com/release/ernie-5-5-flash-base/ · Nov 5, 2026REPObaidu/ERNIE-5.5-Flash-Base model card (Hugging Face)huggingface.co/baidu/ERNIE-5.5-Flash-BaseREPObaidu/ERNIE-5.5-Flash-Base-BF16 model card (Hugging Face)huggingface.co/baidu/ERNIE-5.5-Flash-Base-BF16LABBaidu Research announcement on Xx.com/BaiduResearch/status/2118249227939967424 · Nov 5, 2026LABBaidu ERNIE announcement on Xx.com/Baidu_ERNIE/status/2118342822516437115 · Nov 5, 2026Tracking continues until an independent board lists it.