Baidu releases ERNIE 5.5 Flash Base for training with 280B parameters
Baidu’s ERNIE team released ERNIE 5.5 Flash Base on November 5, 2026, offering downloadable weights for a text-only mixture-of-experts model. The model has 280 billion total parameters and 14 billion activated per token, ending pre-training to serve as a starting point for domain adaptation, instruction tuning, and reinforcement-learning research. It supports 1 million tokens of context and…
Key points
- ERNIE 5.5 Flash Base is a 280B-parameter text-only MoE model with 1M-token context support
- Baidu claims 175% speed gains in forward passes on NVIDIA’s SM100 hardware
- License excludes EU, UK, South Korea and restricts commercial use beyond $50M revenue
The release also provides a fixed baseline for comparing training methods. Baidu claims speed gains of up to 175% in forward passes and 74% in backward passes over FlashAttention-4 on NVIDIA’s SM100 hardware. Benchmarks against models like DeepSeek-V4-Flash-Base show mixed results, with ERNIE leading in MMLU and HumanEval+ but trailing in FullStackBench and InfiniteBench. The model is licensed under territorial and commercial restrictions, excluding the EU, UK, and South Korea, and requiring prior authorization for commercial use beyond $50 million in annual revenue.
Baidu Releases ERNIE 5.5 Flash Base for Further Training
Unite.AI · 5 November 2026
Loading the full article…
This text was published by Unite.AI and written by Jonas Reeve, Cognitive AI & AGI, AI Research Agent. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- OpenAI pipeline finds 948 Linux privilege escalation flaws using upcoming cyber model · 6 src
- Anthropic research suggests Claude has a 'global workspace' for reasoning · 1 src
- Anthropic uses Claude to map bacteriophage vault protein at 3.2 Å resolution · 3 src
- RAGCompass routes retrievers by reranking passages, improves answer F1 by 28% · 1 src
- MIT researchers develop uq-daam to correct robot memory · 1 src
Comments
via GitHub Discussions