Baidu Releases 280B-Parameter ERNIE 5.5 Flash Base
The open-weight mixture-of-experts model is available on Hugging Face and ModelScope under a license that excludes the EU, UK and South Korea.
Updated on Nov. 5, 2026 in Artificial Intelligence
Baidu has released ERNIE 5.5 Flash Base, a text-only mixture-of-experts model with 280 billion total parameters, activating 14 billion per token. The company reports it scored best on 24 of 30 base-model benchmarks in its unified evaluation pipeline.
Why it matters
The release adds a large open-weight model to the field of Chinese and international base models available to developers and researchers outside the company. Its license terms restrict use in three major markets, shaping who can deploy it commercially.
The model has 280B total parameters with 14B active per token, trained on 33.5T tokens using the Muon optimizer, with 43 layers, a 4096 hidden size, and 512 routed experts per MoE layer using top-10 routing plus one shared expert. Its 128K training context extends to 1M for fine-tuning and inference, though the extraction does not report independent verification of these specifications.
The players
Baidu
Baidu is a Chinese technology company operating the Baidu search engine and the ERNIE family of large language models, which it develops through its AI research division.
The details
ERNIE 5.5 Flash Base uses hybrid attention, with Heavily Compressed Attention in 37 layers and Multi-head Latent Attention with sparse attention in six layers. The model is distributed on Hugging Face and ModelScope under a free license for organizations with revenue below US$50 million, excluding the EU, UK and South Korea.
Timeline
• Baidu released ERNIE 5.5 Flash Base on November 5, 2026.
The Tech Race
Baidu's comparison against DeepSeek-V4-Flash-Base places the release directly within the competition among Chinese open-weight base models for benchmark leadership. The mixed result, leading on MMLU-Pro but trailing on HELMET-Rerank, reflects the uneven performance profiles these models show across evaluation suites.
Developers outside the EU, UK and South Korea with annual revenue under US$50 million can download and use the model weights without licensing fees. Organizations in excluded regions or above the revenue threshold will need to review the license terms before deploying it.
The takeaway
Teams evaluating open-weight base models should check license eligibility by region and revenue before building on ERNIE 5.5 Flash Base. Benchmark claims from the model's developer are worth confirming against independent evaluations before relying on them.
Further reading
For more on the models competing in this space, see the Artificial Intelligence section.
