On November 5, 2026, Baidu released the weights for ERNIE 5.5 Flash Base, a text-only AI model. It has 280 billion parameters in total, but only 14 billion are active when processing each token. It also supports contexts of up to 1,048,576 tokens. However, what has been released is the pretrained Base model, without the post-training that improves instruction following and conversational ability. And although few parameters are used at inference time, the model still needs memory for all of its weights, plus memory for a history of up to 1 million tokens. Anyone planning to deploy it needs to weigh whether further training is required, and to check the GPU configuration and memory capacity carefully.
A Base model released as a starting point for further training
ERNIE 5.5 Flash Base is a large language model pretrained from scratch on 33.5 trillion tokens. It is a checkpoint from before mid-training and post-training, and is positioned as a foundation that developers can use to pursue domain adaptation, instruction tuning, reinforcement learning, and other work on their own. Baidu's official announcement also describes the Base model as a starting point for further development.
The weights are distributed through Hugging Face in two versions, FP8 and BF16, in safetensors format. The official usage examples are built around "completions," which continue a given piece of text, rather than the way a tuned chat model would be used out of the box. THE DECODER also points out that developers still have a lot of work ahead of them before the model is ready for practical use.
Baidu plans to release a post-trained "ERNIE 5.5 Flash" as well, but has not given a specific date. For developers who want to adjust a model with their own data and training methods, access to the weights right after pretraining is valuable. For those who want a conversational AI they can use immediately, it is important to distinguish clearly between this Base model and the post-trained version to come.
The mechanisms behind 14B active parameters and long-context processing
The model keeps the parameters used for each token to 14 billion out of 280 billion in total thanks to an architecture called Mixture of Experts (MoE). MoE prepares many specialist networks (experts) and selects only the ones needed for a given input. In ERNIE 5.5 Flash Base, each MoE layer except the first has 512 routed experts, of which 10 are selected and combined with 1 shared expert for processing. Baidu says that using multiple routers and a 2048-dimensional latent space improved expert-selection accuracy and computation and communication efficiency.
Dividing 14 billion by 280 billion gives an active-parameter ratio of 5%. In the official tally, however, the total parameter count includes the input embeddings and the output head, while the active count includes the output head but not the embedding lookups. The 360 million parameters of the additional ECHO module are not included in either figure. The 5% figure is therefore a rough measure of the share of parameters involved in processing each token. It does not mean that computing cost, power consumption, or usage fees are 5% of those of a conventional model.
Handling a context of up to 1 million tokens requires other techniques besides MoE. In large language model inference, information about past tokens is stored in a "KV cache" and reused when processing subsequent tokens. In general, the longer the context, the more memory the KV cache takes up. ERNIE 5.5 Flash Base uses "Heavily Compressed Attention (HCA)" in 37 of its 43 layers, compressing 128 past tokens into a single cache entry to hold a long history efficiently. It also provides a sliding window that can reference the most recent 128 tokens without compression.
The remaining 6 layers use a different approach that combines Multi-head Latent Attention (MLA) and DeepSeek Sparse Attention (DSA). These layers keep per-token information about the past and select important positions to reference. For this reason, even though the HCA layers compress 128 tokens into one, the KV cache of the model as a whole does not shrink to 1/128.
In other words, MoE and HCA are technologies that solve different problems. MoE narrows the parameters used for each token's inference, while HCA reduces the memory needed to retain past context. Combining the two is intended to achieve both computational efficiency for a large model and memory efficiency for long-context processing.
The GPU configuration needed to process 1 million tokens
According to the Hugging Face model card, ERNIE 5.5 Flash Base had a maximum context length of 131,072 tokens during pretraining, and supports up to 1,048,576 tokens in further training and inference. However, the standard launch examples for SGLang and vLLM in the documentation both set 131,072 tokens. Labels such as "128K" and "1M" are approximate token counts and do not refer to character counts. That a model supports 1 million tokens is a separate matter from whether the GPUs at hand can actually process that length.
The vLLM-ERNIE plugin released by Baidu is developed for vLLM 0.32.0 running on NVIDIA GPUs. As a configuration example for handling 1 million tokens, the README lists a parallel setup using four GPUs with 184 GiB of memory each. Even so, fitting the model weights and KV cache is not enough: it explains that the prefill stage, which first processes the long input, needs tens of additional gigabytes of temporary memory.
On the other hand, an FP8 configuration using eight GPUs with 80 GB of memory each can be set to a maximum of 262,144 tokens, and the README states explicitly that 1 million tokens will not fit. This, however, is a condition for a specific configuration shown with the official plugin, not an absolute limit common to all inference implementations. Nor does it give a general minimum number of GPUs needed to run 1 million tokens.
The memory needed to hold the weights also cannot be judged from the 14 billion active parameters alone. The official README gives the weight size as about 280 GB in FP8 and about 560 GB in BF16. When the BF16 weights are spread across four GPUs, the memory left for the KV cache is limited, so the guidance suggests shortening the maximum context length, for example to 65,536 tokens. The key point is that even with 14 billion parameters used at inference, the weights that must be held correspond to the full 280 billion parameters.
The table below organizes the figures given in Baidu's official announcement, the model card, and the vLLM-ERNIE README by what each figure applies to and what constraints remain. It is not a table comparing measured performance.
| Item | What it shows | Caveats and constraints |
|---|---|---|
| 280B total / 14B active parameters | Limits the parameters used for each token to a subset | All weights must still be held; the 5% active ratio is not a reduction rate for cost or power consumption |
| 128 tokens compressed into 1 entry in 37 layers | Reduces the memory needed to store past history in the HCA layers | Because of the uncompressed reference window and the 6 layers using a different method, the overall KV cache does not shrink to 1/128 |
| Up to 1,048,576 tokens / standard launch example of 131,072 tokens | The model's supported limit versus the context length actually specified at launch | Using the full supported limit requires separately checking GPU memory capacity and settings |
| Up to 262,144 tokens in an FP8 configuration with 8× 80GB GPUs | Operating condition for a specific configuration shown in the official vLLM implementation | 1 million tokens will not fit; even in other supported configurations, extra memory is needed for prefill |
In the model name and specifications, "14B" means 14 billion and "280B" means 280 billion parameters. The 14B/280B ratio is the share of active parameters, 128 is the unit of history compression in the HCA layers, and 1M is the upper limit of the supported context length. None of these figures directly indicates the memory usage or operating cost of an actual system as a whole.
When actually deploying the model, you first need to confirm that all the weights fit on the GPUs, then verify that enough memory remains to handle the required context length. Inference speed and operating cost depend on the GPU configuration, the number of concurrent requests, input length, and other factors, so measurement in the real environment is essential. The specification that the model supports 1 million tokens alone does not tell you the response quality, processing speed, or power consumption when handling that length.
Comparison with similarly sized Base models shows strengths and weaknesses
To verify the performance of ERNIE 5.5 Flash Base, Baidu also published comparison results against publicly available Base models of similar scale. The comparison targets include "DeepSeek V4 Flash Base," "Hy3 preview Base," and "MiMo V2.5 Base." Baidu explains that it used a common test environment, prompts, and few-shot examples, and judged open-ended questions with the same evaluation model. All comparisons are between Base checkpoints, not a ranking that mixes in post-trained chat models.
According to the published evaluation results, in the comparison with DeepSeek V4 Flash Base, ERNIE 5.5 Flash Base came out ahead on MMLU 5-shot exact match (93.3 vs. 89.4) and HumanEval+ 1-shot Pass@1 (82.3 vs. 69.5). This shows strengths in some evaluations of knowledge questions and code generation. On the other hand, it fell behind DeepSeek V4 Flash Base on ChineseSimpleQA 10-shot (73.0 vs. 75.3) and HELMET-Rerank 0-shot NDCG@10 (56.2 vs. 60.0). It is not superior on every evaluation item, which suggests that strengths differ by use case.
These results can inform the choice of a base model for further training. However, they are benchmarks at the pretraining stage published by Baidu, and do not directly show performance after tuning for a specific purpose, or practicality as a finished chat model. Nor do they guarantee the processing speed or response quality when actually using a 1-million-token context. Developers considering adoption need to choose evaluation metrics that fit their use case and validate the further-trained model under real usage conditions.
Commercial use is restricted by revenue and region
The fact that the model weights are public does not mean there are no restrictions on their use. ERNIE 5.5 Flash Base is governed by a custom license agreement v1.0, with different conditions for non-commercial and commercial use. Non-commercial use incurs no model license fee. Commercial use is also free of model license fees if the annual revenue of the related products and services is US$50 million or less. However, a "Powered by ERNIE" notice and a one-time notification before commercial use begins are required. If annual revenue exceeds US$50 million, a separate written commercial-use license must be obtained in advance.
Annual revenue here refers to revenue for the most recent 12 months up to the time of determination, not a calendar or fiscal year figure. Revenue is also counted across the related products and services as a whole, even if they use other AI models or technologies alongside ERNIE. It is not possible to carve out only the revenue directly attributable to ERNIE for the determination. In addition, the license includes provisions that aggregate the revenue of companies under common control, so caution is needed when using the model through multiple legal entities.
There are also regional restrictions. The license excludes the European Union (EU), the United Kingdom, and South Korea from the licensed territory, and does not permit the use or distribution of the model, derivative models, or outputs in those regions. It also prohibits using ERNIE's weights, derivative models, or generated outputs to improve AI models or machine learning systems other than ERNIE and its derivatives. Even though modifying and further training the model are permitted, the license terms continue to apply to the use and distribution of derivative models.
ERNIE 5.5 Flash Base is therefore best understood as an "open-weight" model: one whose weights you can obtain and use or modify on your own. Especially when considering embedding it in a commercial service, you need to check not only the model's performance and GPU requirements but also how revenue is calculated, the regions served, the attribution requirement, and the treatment of derivative models.
ERNIE 5.5 Flash Base is characterized by a design that aims to use a very large model of 280 billion parameters efficiently, through an MoE structure with 14 billion active parameters and KV cache compression. However, high computational efficiency does not necessarily mean the model can run on little GPU memory, and supporting 1 million tokens also requires substantial hardware. For deployment, it is important to satisfy the license terms, secure a GPU configuration that can handle the required context length, and verify performance after further training. Whether you can put a development setup in place for that will decide whether to adopt this Base model or wait for the post-trained version, which has no release date yet.