ERNIE-5.5-Flash-Base
Baidu's open-weight 280B MoE base model (14B active), pretrained on 33.5T tokens, with 1M-token context at inference.
Overview
ERNIE-5.5-Flash-Base is a text-only mixture-of-experts base checkpoint from Baidu, listed on Hugging Face with a November 5, 2026 announcement date. It has 280B total parameters with 14B active per token, 43 layers and 512 routed experts (top-10) plus one shared expert, and was pre-trained on 33.5T tokens up to 128K context; fine-tuning and inference support 1M tokens. It is the end-of-pretraining checkpoint, before mid- or post-training, with FP8 and BF16 weights, released under the ERNIE 5.5 Flash Base Model License Agreement v1.0.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- 280B total / 14B active parameters (Safetensors metadata lists 286B)
- 1M-token context for fine-tuning and inference; 128K in pretraining
- Heavily Compressed Attention and Multi-head Latent Attention layers
- Reported MMLU 93.3, MMLU-Pro 78.4, HumanEval+ 82.3 versus DeepSeek-V4-Flash-Base and others
- FP8 and BF16 weights under a custom Baidu license
Use cases
- Starting checkpoint for continued pretraining or domain fine-tuning
- Research on sparse-attention MoE architectures
- Base-model comparison against DeepSeek-V4-Flash-Base
Pricing
Free weights on Hugging Face under the ERNIE 5.5 Flash Base Model License Agreement v1.0; no hosted pricing published.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 5 Nov 2026
- Record updated
- 6 Nov 2026
- Record last checked
- 6 Nov 2026