Chinese technology company Baidu has released ERNIE 5.5 Flash Base, a new open-weight artificial intelligence model designed for advanced reasoning, coding, long-context processing and further developer training. Announced on November 5, 2026, the model contains 280 billion total parameters but activates only 14 billion parameters per token, using a sparse Mixture-of-Experts (MoE) architecture to balance computing efficiency with model capacity.
The release is based on 33.5 trillion training tokens and supports a context window of up to one million tokens for fine-tuning and inference. Baidu has published the model weights in FP8 and BF16 formats, giving developers access to its underlying architecture and pretrained capabilities. The company says the model performs competitively against similarly sized open-weight models, including offerings from DeepSeek and other Chinese AI developers. A post-trained ERNIE 5.5 Flash version is expected to follow.
What Is Baidu’s ERNIE 5.5 Flash Base?
ERNIE 5.5 Flash Base is a large language model developed by Baidu’s ERNIE team. It is designed as a foundation for developers who want to build specialised AI applications rather than simply use a finished conversational assistant.
The model has 280 billion parameters in total, with only 14 billion activated for each token processed. This architecture allows it to retain the capacity of a much larger model while limiting the amount of computation required for individual operations.
Its key specifications include:
| Specification | ERNIE 5.5 Flash Base |
|---|---|
| Developer | Baidu |
| Total parameters | 280 billion |
| Active parameters | 14 billion |
| Architecture | Sparse Mixture of Experts |
| Pre-training data | 33.5 trillion tokens |
| Maximum training context | 128K tokens |
| Fine-tuning and inference context | Up to 1 million tokens |
| Available formats | FP8 and BF16 |
| Model type | Pretrained base model |
The model is particularly designed for developers working on reasoning systems, coding applications, long-document analysis and autonomous AI agents.
How Mixture-of-Experts Architecture Improves Efficiency
One of the main features of ERNIE 5.5 Flash Base is its sparse Mixture-of-Experts architecture.
Traditional dense AI models generally activate their entire parameter network for each token. This can make inference expensive, especially when models contain hundreds of billions of parameters.
MoE models take a different approach. They divide parts of the network into specialised components called experts. A routing system selects which experts should process a particular input.
ERNIE 5.5 Flash Base contains 280 billion total parameters but activates only 14 billion per token. This means that a relatively small portion of its total network is involved in processing each token.
The model uses 512 routed experts in its MoE layers, selecting 10 routed experts alongside one shared expert.
This architecture is intended to reduce computation and communication costs while retaining substantial model capacity.
However, activating only 14 billion parameters does not mean the model can run on hardware designed for a 14-billion-parameter dense model. Its full weights and supporting infrastructure still create significant memory and deployment requirements.
Baidu Trains the Model on 33.5 Trillion Tokens
Baidu says ERNIE 5.5 Flash Base was pretrained from scratch on 33.5 trillion tokens.
The training process included multilingual web information, software repositories, programming data and long documents. The company placed particular emphasis on capabilities that are difficult to develop through later training, including repository-level coding and understanding lengthy information.
The model’s training context was gradually extended to 128,000 tokens.
For fine-tuning and inference, however, Baidu supports context lengths of up to one million tokens. This allows developers to experiment with applications involving very large documents, extensive codebases and long-running information-processing tasks.
A one-million-token context window could be useful for analysing large collections of documents without repeatedly dividing them into smaller inputs.
It could also support AI agents that need to retain substantial amounts of information while completing multi-step tasks.
ERNIE 5.5 Flash Base Shows Strong Benchmark Results
Baidu published benchmark comparisons against three similarly sized models: DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base.
The company said all four models were tested using the same evaluation framework, prompts and few-shot settings. Its published results show ERNIE 5.5 Flash Base performing strongly across several reasoning, knowledge, mathematics and coding evaluations.
Selected results include:
| Benchmark | ERNIE 5.5 Flash Base | DeepSeek-V4-Flash-Base |
|---|---|---|
| MMLU | 93.3 | 89.4 |
| MMLU-Pro | 78.4 | 69.2 |
| GPQA | 60.2 | 55.9 |
| HumanEval+ | 82.3 | 69.5 |
| MBPP+ | 81.5 | 67.2 |
| LiveCodeBench | 40.1 | 31.9 |
| LongBench-v2 | 48.7 | 44.5 |
The results indicate that Baidu’s model has competitive performance across several categories. However, it does not lead every benchmark.
For example, Hy3-preview-Base achieved a higher result on FullStackBench, while DeepSeek-V4-Flash-Base performed better on ChineseSimpleQA and InfiniteBench.
These comparisons are also based on evaluations published by Baidu. Independent testing will be important to establish how the model performs across different hardware environments and real-world applications.
Why the One-Million-Token Context Window Matters
Long-context processing has become an important area of competition among AI developers.
A context window determines how much information a model can process within a single interaction. Larger context windows can help AI systems work with lengthy reports, books, software repositories, legal documents and extensive research material.
ERNIE 5.5 Flash Base supports up to one million tokens during fine-tuning and inference, potentially allowing developers to build applications that process substantial amounts of information.
For businesses, this could be useful in financial research, document analysis, customer support systems and enterprise knowledge management.
For software developers, the ability to process large codebases could support repository-level analysis, debugging and code modification.
However, a larger context window does not automatically guarantee accurate retrieval or reasoning across every part of a long document. Performance also depends on the model’s architecture, attention mechanisms and quality of the information supplied.
Baidu Makes Model Weights Available to Developers
Baidu has released ERNIE 5.5 Flash Base weights through Hugging Face in two formats:
- FP8: A lower-precision format designed to reduce memory requirements.
- BF16: A higher-precision format commonly used for large-scale AI training and inference.
The model is available under Baidu’s ERNIE 5.5 Flash Base Model License Agreement version 1.0. Developers should review the licence conditions before deploying it, particularly for commercial applications.
The release is aimed at developers and research teams that want to customise the model through domain adaptation, instruction tuning and reinforcement learning.
Baidu has also provided instructions for running the model using tools such as Transformers, SGLang and vLLM.
An important distinction is that the released checkpoint has not undergone mid-training or post-training. It is therefore not equivalent to a finished chatbot that is already optimised for following user instructions.
Baidu says a post-trained ERNIE 5.5 Flash model will be released subsequently.
China’s AI Competition Moves Further Into Open-Weight Models
Baidu’s release comes amid intensifying competition among Chinese AI developers, including DeepSeek, Moonshot AI and other companies developing large language models.
Chinese AI companies have increasingly focused on making model weights available to developers, encouraging experimentation, fine-tuning and deployment across different applications.
This approach can help companies expand their developer ecosystems while creating alternatives to proprietary AI services.
The commercial importance of this strategy was also highlighted in a November 6 report by the South China Morning Post. Citing research from Menlo Ventures, the publication reported that Chinese AI models accounted for 89% of estimated global enterprise spending on open-source models, with Moonshot AI’s Kimi leading the category.
That figure refers specifically to spending on open-source models, not the entire enterprise AI market. Proprietary AI systems continue to account for the much larger share of overall enterprise AI spending.
The Bigger Picture
ERNIE 5.5 Flash Base demonstrates how Chinese AI developers are competing through a combination of large-scale model training, sparse architectures and developer access to pretrained weights. Its 280-billion-parameter design, 14-billion active parameter count and one-million-token inference context highlight the industry’s focus on making powerful models more efficient and adaptable.
For Baidu, releasing the base model could strengthen its position among developers building customised AI systems. For the wider AI industry, it adds another major model to an increasingly competitive open-weight ecosystem. The commercial impact will depend on independent performance evaluations, deployment costs, licensing conditions and the quality of the post-trained version.
Looking Ahead
The next major development will be the arrival of Baidu’s post-trained ERNIE 5.5 Flash model. Its performance will help determine whether the underlying capabilities of the base checkpoint translate into a practical assistant for coding, reasoning, research and business applications.
Developers will also be watching how efficiently ERNIE 5.5 Flash Base can be fine-tuned and deployed across different hardware configurations. If its long-context architecture and sparse computation deliver meaningful cost advantages, Baidu could attract more developers seeking alternatives to proprietary AI platforms and other leading open-weight models.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
