Base model released ahead of the chat version
China's Baidu released the weights (trained parameters) of its AI model "ERNIE-5.5-Flash-Base" on Hugging Face on Nov. 5. It is the model as of the end of pretraining, without additional training to make it follow instructions. Baidu positions it as a foundation on which companies and researchers can do further training for their own fields, and said the chat-oriented "ERNIE 5.5 Flash" will come later. It did not give a timeline.
Baidu's ERNIE-5.5-Flash-Base weights just went public on Hugging Face, before any announcement
@aitrackerbot, an X account that tracks new model releases, reported that the weights were released before an official announcement. This is the account's own assessment, and no announcement on Baidu's official X account has been confirmed. Baidu's technical blog is dated Nov. 5, but whether it came before or after the post has not been confirmed.
The model is a "mixture of experts (MoE)" type that uses only some of many expert components, with 280 billion total parameters, 14 billion of which are used to process each token. It was trained from scratch on 33.5 trillion tokens. Its context window is up to 128,000 tokens during training, and it supports 1 million tokens for fine-tuning and inference. It handles text only. The tally on the Hugging Face page shows about 286.2 billion total parameters.
EU, Britain and South Korea excluded; revenue conditions for commercial use
Anyone can download the weights, but the license is Baidu's own. It covers the entire world except the EU, Britain and South Korea, where neither the model nor derivative models built through further training nor its outputs may be used. Japan is not among the excluded regions.
For commercial use, no license fee is charged if the annual revenue of related products or services is $50 million or less. However, users must notify Baidu before use and display "Powered by ERNIE." Above $50 million, written permission from Baidu is required. Revenue is combined across companies under the same control. Using the model or its outputs to improve other AI is also prohibited.
For "ERNIE 4.5," released in 2025, Baidu adopted the less restrictive Apache 2.0 license. "Unlimited-OCR-2," a document-reading model whose weights it released on Oct. 27, is also under the MIT license. Among the models it was compared against, Xiaomi's MiMo-V2.5 is under the MIT license, and DeepSeek-V4-Flash-Base is also reported to be under the MIT license.
Tops 24 of 30 benchmarks in its own comparison
Baidu compared the model with DeepSeek-V4-Flash-Base, China's Tencent's "Hy3-preview-Base" and Xiaomi's "MiMo-V2.5-Base" on 30 benchmarks covering knowledge, math, programming, long text and more. The three models are of similar scale, with 284 billion to 310 billion total parameters, and all four are at the pretraining stage. In results Baidu measured with the same evaluation environment and the same prompts, ERNIE-5.5-Flash-Base ranked first on 24 benchmarks. All figures are on a scale with a maximum of 100 and were announced by Baidu on Nov. 5.
The gap was widest on MMLU-Pro, which tests broad knowledge, where it scored 78.4, ahead of DeepSeek-V4-Flash-Base's 69.2. On the programming benchmark LiveCodeBench, it also scored 40.1, ahead of Hy3-preview-Base's 33.5.
Others lead on six benchmarks, including long text
On the remaining six benchmarks, other companies' models came out ahead. On InfiniteBench, which tests reading long texts, it scored 46.5, short of DeepSeek-V4-Flash-Base's 48.0 and MiMo-V2.5-Base's 46.7. On ChineseSimpleQA, which tests knowledge in Chinese, its 73.0 was also below DeepSeek-V4-Flash-Base's 75.3.
| Benchmark | ERNIE-5.5-Flash-Base | DeepSeek-V4-Flash-Base | Hy3-preview-Base | MiMo-V2.5-Base |
|---|---|---|---|---|
| ChineseSimpleQA | 73.0 | 75.3 | 73.2 | 68.8 |
| BBH (CoT) | 89.2 | 87.5 | 90.0 | 89.8 |
| INCLUDE-lite | 79.5 | 79.8 | 77.7 | 74.5 |
| FullStackBench | 54.6 | 43.3 | 56.7 | 54.1 |
| HELMET-Rerank | 56.2 | 60.0 | 55.7 | 56.6 |
| InfiniteBench | 46.5 | 48.0 | 33.0 | 46.7 |
Rivals compared are models from around April; no third-party evaluation found
The three models compared were all released around April. The companies have since released newer models: DeepSeek released V4.1 Flash in September and improved it in October, and Tencent released the official version of "Hy4" on Nov. 4. Neither is included in Baidu's comparison. As far as ORIENS checked, no third-party evaluation under the same conditions has been published. Artificial Analysis also did not have a page for this model as of Nov. 5.
Running it takes multiple GPUs; next milestones on the 12th and 16th
The model is about 280 gigabytes in its FP8 version and about 560 gigabytes in its BF16 version. In the guide for the inference software SGLang, the recommended configuration for the FP8 version is eight Nvidia H100s or four B200s. It does not support chat and works by generating a continuation of text.
Baidu will hold its annual event "Baidu World 2026" in Beijing on Nov. 12 and report July-September quarter results before the U.S. market opens on the 16th. The focus is whether a timeline for the chat-oriented ERNIE 5.5 Flash will be given. The SGLang guide lists that model as "coming soon."
