baidu/ERNIE-5.5-Flash-Base
AI model · https://huggingface.co/baidu/ERNIE-5.5-Flash-Base
First seen by Oriel on .
AI summary
ERNIE-5.5-Flash-Base is Baidu’s 280B-parameter sparse MoE base language model, with 14B activated parameters and support for up to 1M-token fine-tuning and inference.
- The pre-training checkpoint was trained from scratch on 33.5T tokens with a maximum training context of 128K, and is released for domain adaptation, instruction tuning, and reinforcement-learning research.
- Its 43-layer architecture combines 37 Heavily Compressed Attention layers with six MLA-DSA layers, plus 512 routed experts per MoE layer and one shared expert.
- The model introduces Multi-Router Latent MoE, ECHO future-token prediction, manifold-constrained Hyper-Connections, and a MultiMax output head.
- On the listed unified benchmarks, it leads the compared models on MMLU-Pro, GPQA, HumanEval+, LiveCodeBench, and LongBench-v2, but trails on some measures including HELMET-Rerank and InfiniteBench.
- FP8 and BF16 safetensors are provided, with production serving instructions for SGLang and vLLM; the model is under the ERNIE 5.5 Flash Base Model License Agreement v1.0.
#large language models #mixture of experts #long context #model serving
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Likes
Total likes reported by the source.
- Hugging Face Models: 353, latest observation ; +180 over observed period since
Downloads (30d)
Rolling 30-day downloads. A change in this value is not new downloads during the observed period.
- Hugging Face Models: 4,145, latest observation ; +4,132 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Models | baidu/ERNIE-5.5-Flash-Base | https://huggingface.co/baidu/ERNIE-5.5-Flash-Base Read discussion |