Baidu Promises Apache 2.0 Weights for ERNIE 5.5 Flash by November 25. Today It Is API-Only, and the Base Weights It Already Released Cannot Be Used in the EU, the UK, or South Korea.
Baidu released ERNIE 5.5 Flash Preview on November 11, the first post-trained model on the ERNIE 5.5 Flash Base it published November 5: 280 billion total parameters, 14 billion active, a 1 million token context, on the Qianfan API now, with weights to be open-sourced under Apache 2.0 soon per the English page and by November 25 per Chinese press. Baidu calls it an intermediate checkpoint whose capabilities may still fall short. The Base weights on Hugging Face carry Baidu's own license, whose territory is worldwide except the EU, the UK, and South Korea, so the post-trained model would be more open than its own base if the Apache promise holds. Benchmarks are Baidu's charts, some peer scores self-evaluated, and there is no independent score. We read the release pages and the license.
By T.W. Ghost
Baidu released ERNIE 5.5 Flash Preview today, November 11, 2026. It is the first post-trained model on the ERNIE 5.5 Flash Base that Baidu put on Hugging Face on November 5: a 280-billion-parameter mixture of experts with 14 billion active, a 1 million token context, available now on the Qianfan API. The English release page says the weights "will be open-sourced under Apache 2.0 soon"; Chinese materials, relayed by press, give November 25. Baidu's own framing is modest: the Preview is "an intermediate checkpoint from post-training," training is not finished, and "some capabilities may still fall short."
The licensing is the part to read twice. The Preview is API-only today and promised as Apache 2.0. The Base it was trained from is already downloadable, under Baidu's own license, whose definition of "Territory" is worldwide other than the European Union, the United Kingdom, and South Korea, with use, distribution, and even outputs outside that territory not authorized, revenue-based commercial conditions, and Hong Kong law and courts. If the Preview ships as Apache 2.0 on schedule, the post-trained model will be more open than its own base, and a European lab will be allowed to run the fine-tune but not the thing it was fine-tuned from. We read the Preview and Base release pages, the Hugging Face card, the license file, and the press relays. Here is what held up.
What Baidu Shipped, and Under What Terms
| ERNIE 5.5 Flash Preview (November 11, 2026) | |
|---|---|
| Architecture | 280B total, 14B active MoE; 1M context |
| Access today | Qianfan API only |
| Weights | "Open-sourced under Apache 2.0 soon" (English page); by November 25 (press relay of Chinese announcement); not on Hugging Face as of this reading |
| Baidu's framing | Intermediate post-training checkpoint; some capabilities may still fall short |
| Base model | ERNIE 5.5 Flash Base, November 5; 33.5T pretraining tokens; 128K training context; FP8 and BF16 safetensors on Hugging Face |
| Base license | Baidu's own: Territory is worldwide except the EU, UK, and South Korea; use, distribution, and outputs outside it not authorized; revenue-based commercial conditions; Hong Kong law |
| Qianfan promo | 1 yuan in, 4 yuan out per 1M at peak, half list (press; not checked on the price page) |
| Speed claim | 200-plus tokens per second on Qianfan (vendor, via X summary; unverified) |
| Independent score | None |
A date is not a file. Until a repository with the Preview's weights and a LICENSE that says Apache 2.0 exists, the model is an API with a promise attached, the same position Hy4 and Inkling 2 Small were in last week.
Two licenses, one lineage. A European or Korean team cannot legally use the Base today. Whether they can use an Apache 2.0 Preview trained from it is a question for their lawyers, and it is a question Baidu's release pages do not address.
Baidu grades Baidu. The benchmark charts are images on the release page, not tables we transcribed, and press notes that some peer scores in them are Baidu's own evaluations of other labs' models. The summaries say the Preview trails peers on Terminal-Bench 4.0 and leads on AutomationBench and SpreadsheetBench V2, with figures of 46.5 and 61.9 circulating from an X summary. We print them as what they are: a vendor chart read secondhand.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Released Nov 11; 280B / 14B active; 1M context; Qianfan API | Vendor (release page) |
| Apache 2.0 weights "soon" | Verified wording (English release page) |
| Weights by November 25 | Press relay of Baidu's Chinese announcement |
| Preview weights not on Hugging Face; Base and BF16 are the newest Baidu uploads | Verified (Hugging Face) |
| Intermediate checkpoint; capabilities may fall short | Verified (Baidu's own text) |
| Base license territory excludes EU, UK, South Korea; outputs outside not authorized | Verified (LICENSE file) |
| Terminal-Bench 4.0 46.5; AutomationBench 61.9 | Vendor chart via X summary; not transcribed by us |
| 200-plus tokens per second; Qianfan promo price | Vendor and press; unverified |
| Demos: 11-hour Unreal Engine wuxia game, mini-games, Blender airport, FreeCAD engine | Vendor demos |
| Artificial Analysis score | None |
Who Should Use It
Watch Hugging Face on November 25, and read the LICENSE file that lands with the weights before you read the model card; the word Apache in a blog post is not the license on the repo.
Skip it today if you are in the EU, UK, or South Korea and were planning to build on the Base, because Baidu's license says you may not, and skip the chart until a board Baidu does not run scores the model.
The frontier is elsewhere: in the same size class, DeepSeek V4.1 Flash 1017 has MIT weights you can download today and an independent score. Our model comparison has the current standings side by side.
Sources
- Baidu ERNIE: ERNIE 5.5 Flash Preview release
- Baidu ERNIE: ERNIE 5.5 Flash Base release
- Hugging Face: baidu/ERNIE-5.5-Flash-Base
- Sulat: Baidu ERNIE 5.5 Flash Preview, weights by 25 November (press)
- KuCoin News: Baidu ERNIE 5.5 Flash Preview released (press)
*Not sure whether a promised Apache release or a restricted base model fits your jurisdiction? Take the free 2-minute quiz and get matched. Then read about the other Chinese flagship that appeared the same morning with no numbers and a two-week life: Qwen4-Max Preview.*