Baidu opens ERNIE 5.5 Flash Base to developers
Baidu releases ERNIE 5.5 Flash Base, a 280-billion-parameter open-source foundation model. The version comes without fine-tuning, designed for reasoning and long-context use cases.
Baidu has released ERNIE-5.5-Flash-Base, an open foundation model that deliberately comes without fine-tuning or post-training. Introduced on November 5, 2026, it leaves developers in control of adapting the system, with a post-trained version announced to follow. The model has 280 billion parameters, including 14 billion active per query, a format that places it in the race for large open models for reasoning, coding, and long context sequences.
An open model, but not ready to use
Baidu does not deliver a finished assistant here, but a working base. ERNIE 5.5 Flash Base is described as an open source foundation model without final tuning, meaning the company provides the weights and architecture, but not a polished application for the general public. The weights are available on Hugging Face in FP8 and BF16, two common numeric formats to reduce memory needed for inference and training.
In this context, the strategy is quite clear: to push adoption among technical teams who want to control their own specialization. It is less flashy than a shiny demo, but often more useful. "Plug-and-play" models make the coffee; raw models sell the espresso machine.
The technical promise mainly lies in scale
ERNIE 5.5 was pre-trained on 33.5 trillion tokens, with a training context window up to 128,000 tokens. Baidu indicates, however, that the model is designed to support up to one million tokens in fine-tuning and inference. The mixture-of-experts allows activating only part of the network per query, here 14 billion parameters out of 280 billion total. Result: the displayed size remains massive, but the computation cost per input can be better controlled.
On benchmarks cited by Baidu, ERNIE 5.5 Flash Base competes with open models of comparable size such as DeepSeek-V4-Flash-Base, Hy3-preview-Base, and MiMo-V2.5-Base. The results cover several domains, including reasoning, mathematics, coding, multilingualism, and long context. Baidu claims its version is in the same competitive category, without publishing a detailed table in the provided source [to be verified].
What this release changes for the market
The release fits into a broader battle around large open models, where value no longer comes only from raw size but from ease of adoption by developers. An open model with large token volumes, an extended context window, and publicly available weights can serve as a base for specialized services, agents, or internal tools. For Chinese players as well as their international competitors, the issue is no longer just releasing a model, but keeping developers engaged over time.
At this stage, Baidu also announces that a post-trained version will follow. The real question will therefore be less whether the model exists, but who adopts it, how quickly, and for which concrete uses.
Key points
- Released on November 5, 2026, according to Baidu source.
- 280 billion parameters, including 14 billion active.
- 33.5 trillion pre-training tokens.
- Context window announced up to one million tokens.
- Weights available on Hugging Face in FP8 and BF16.
- Baidu targets DeepSeek-V4-Flash-Base and MiMo-V2.5-Base.
By the numbers
- 280 billion — total parameters of ERNIE 5.5 Flash Base, Baidu source, 2026.
- 14 billion — active parameters per input, Baidu source, 2026.
- 33.5 trillion — pre-training tokens, Baidu source, 2026.
- 128,000 tokens — announced training context, Baidu source, 2026.
- 1 million tokens — targeted context in fine-tuning and inference, Baidu source, 2026.