Further training on the base model, offered on Qianfan and OpenRouter
China's Baidu on Nov. 11 announced the AI model "ERNIE 5.5 Flash Preview." It is the first version built on "ERNIE-5.5-Flash-Base," whose weights were released on Nov. 5, with additional training to follow instructions (post-training). Its API is now offered on Baidu's cloud "Qianfan," and the company said it will soon release the weights under the Apache 2.0 license.
It was also listed on Nov. 10 (Nov. 11 Japan time) on OpenRouter, which offers multiple AI models through a single API. Serving is handled by Baidu's Qianfan, and the version available on OpenRouter is a free version with no charges. Baidu's announcement page also links to this free version on OpenRouter.
@aitrackerbot, an X account that automatically detects and posts listings of new models, reported the listing on Nov. 10 (Nov. 11 Japan time). ORIENS has not been able to confirm an announcement on Baidu's official X account.
The architecture is the same as the base model released on Nov. 5. It is a "mixture of experts (MoE)" model that uses only some of many expert components, with 280 billion total parameters, of which 14 billion are used to process each token. It can handle a context of 1 million tokens at a time and works with text only.
Free version supports reasoning and tool calls; paid version pricing not listed
The free version on OpenRouter can read 1,024,000 tokens at a time and output up to 921,600 tokens in a single response. Supported parameters include setting the depth of reasoning (reasoning_effort), calling external tools and programs (tools) and answering in a fixed format (structured_outputs).
OpenRouter also has a listing for a paid version, but as of Nov. 11 no provider was registered and no pricing was shown. ORIENS has not been able to confirm pricing on Qianfan.
First on 2 of 8 items in company's own comparison
Baidu compared the model with 6 others, including Anthropic's "Claude Opus 5.5," OpenAI's "GPT-6.1 Sol," China's Tencent's "Hy4," DeepSeek's "V4.1 Flash 1017" and China's Zhipu's "GLM-5.3," on 8 items measuring the ability of AI to carry out work on its own, office tasks and more. In the results Baidu presented on Nov. 11, ERNIE 5.5 Flash Preview ranked first on 2 items. On AutomationBench, which measures business automation, it scored 61.9, ahead of second-place Claude Opus 5.5 at 58.0. On SpreadsheetBench V2, which measures spreadsheet work, it scored 36.4, ahead of Claude Opus 5.5 at 35.1.
The chart does not state maximum scores. Many of the other companies' figures carry a mark indicating that they are self-reported scores. As far as ORIENS checked, no third-party evaluation under the same conditions has been published.
Terminal evaluation trails Claude Opus 5.5 by 17.1 points
On the remaining 6 items, other companies' models scored higher. On Terminal-Bench 4.0, which measures the ability to operate a computer through command input, it scored 46.5, short of Claude Opus 5.5 at 63.6 and GPT-6.1 Sol at 56.1. On DeepSWE v1.1, which measures fixing software bugs, it scored 71.2, below first-place DeepSeek V4.1 Flash 1017 at 77.0, ranking fifth of 7 models. On APEX-Agents 1.1, which measures professional work, it scored 64.2, short of Hy4 at 75.0 and Claude Opus 5.5 at 72.9.
Baidu describes this version as being at an intermediate stage of post-training. While it said the model is competitive with models of similar size, it said some capabilities may be lacking because training is not finished.
Weights to be released under Apache 2.0, on looser terms than the base model
Baidu said it will soon release the weights of ERNIE 5.5 Flash Preview under the Apache 2.0 license. It did not give a timeline.
The base model released on Nov. 5 is under Baidu's proprietary license, cannot be used in the EU, Britain or South Korea, and requires Baidu's written permission for commercial use when annual revenue from related products exceeds $50 million. Apache 2.0 sets no restrictions based on region or revenue. If things go as announced, the chat-oriented version will be usable on broader terms than the base model.
As examples, Baidu cited writing a game running on the game engine "Unreal Engine 5.8" in C++ in a single 11-hour session, building structures in the 3D production software "Blender," and creating spreadsheets and documents.
Another model, "ERNIE-Light-DT-1.0," listed the same day
Baidu's "ERNIE-Light-DT-1.0" was also listed on OpenRouter on the same day, Nov. 10 (Nov. 11 Japan time). Pricing is $1.13 for input and $3.94 for output per million tokens, and $0.113 when reusing input that has already been read. It can read 1,024,000 tokens at a time and output up to 131,072 tokens in a single response.
According to OpenRouter's description, it is built on a publicly released GLM model and incorporates research results from models of Baidu's search division. It uses "Baidu Sparse Attention," Baidu's own attention mechanism that limits computation to a subset. Baidu's announcement page does not mention this model.
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 Nov. 16. The focus will be on whether it gives timing for the official version of ERNIE 5.5 Flash and the release of its weights.
