On this page
Baidu's English product-updates page for this release carries the calendar date November 11, 2026, under the kicker "Product updates / November 11, 2026." The Chinese announcement page carries 「产品动态 / 2026年11月11日」. On that calendar day, Baidu released ERNIE 5.5 Flash Preview, calling it the first post-trained model built on the ERNIE 5.5 Flash Base released earlier.
The official announcement uses the spaced display name ERNIE 5.5 Flash Preview. Baidu Qianfan lists the same release as ERNIE-5.5-Flash-Preview, with the formal ID ernie-5.5-flash-preview. Both forms appear in official materials.
The model has 280 billion total parameters, 14 billion active parameters, and a 1M-token context. The English announcement says it is strong at coding, 3D modeling, knowledge work, and content creation, that it balances capability with inference efficiency, and that the API is now available on the Qianfan platform. Those lines are official self-description, not independently measured scores. Official wording also describes ERNIE 5.5 Flash Preview as an intermediate checkpoint from post-training: it is already competitive with models of its size, but because training is not finished, some capabilities may still fall short. Post-training is ongoing, and Baidu says improvements will follow in the official release. This Preview is a post-trained intermediate version now offered through the Qianfan API, not a generally available launch and not a Flash Base weights update.
The existing article What It Takes to Turn ERNIE-5.5-Flash-Base Into a Useful Assistant is about turning Flash Base into a useful assistant. That coverage is separate from this Preview API release. The official ERNIE homepage lists Flash Base as a product update dated November 5, 2026; the Preview announcement page itself only says the Base model was "released earlier."
How Qianfan lists the model
Qianfan's public model list, which shows a document date of 2026-11-10, places the same formal ID, ernie-5.5-flash-preview, under both text generation (文本生成) and deep thinking (深度思考), in the ERNIE flagship-model rows of each category. In both listings, context length, maximum input, and maximum output are all 1M. Default rate limits are RPM=60 and TPM=1,000,000 (the page also writes 1000000).
For deep thinking, the documentation includes this official note: 「注：深度思考模型在输出最终回答之前会先输出一段思维链内容，以下最大输出长度为包含思维链输出的总长度。」 The 1M maximum output is the total length including the thinking chain, not a cap that excludes it.
The announcement pages include an Access API / 获取API control that points to the Qianfan console. An unauthenticated visit shows only a login shell. The formal ID, rate limits, and prices below are taken from Qianfan's public documentation.
Official pay-as-you-go prices
Qianfan's billing documentation, also marked 更新时间：2026-11-10, publishes pay-as-you-go prices in yuan per thousand tokens (元／千tokens). The tables for text generation and deep thinking are the same. The model name is listed as ERNIE-5.5-Flash, the version name as ERNIE-5.5-Flash-Preview, and the service as inference (推理服务). Batch inference is marked "-" in the table.
For peak hours, 8:00–22:00, the official figures are:
- Input: promotional rate 0.001, list rate 0.002
- Cache hit: promotional rate 0.0001, list rate 0.0002
- Output: promotional rate 0.004, list rate 0.008
For idle hours, 22:00–next day 8:00, the official figures are:
- Input: promotional rate 0.0005, list rate 0.001
- Cache hit: promotional rate 0.00005, list rate 0.0001
- Output: promotional rate 0.002, list rate 0.004
The price table does not label a timezone. The documentation does not state an end date for the promotional rates. These are paid pay-as-you-go prices, not a free allowance. The 2026-11-10 stamp is the Qianfan documentation date; it is not the same calendar day as the November 11 announcement.
Open-source wording, which is not the same in English and Chinese
Official English and official Chinese do not use one combined open-source sentence.
The English announcement states: "The weights will be open-sourced under Apache 2.0 soon."
The Chinese announcement states: 「权重将于11月25日前开源」. Its meta description says 「模型权重将于11月25日前开源。」 The Chinese announcement does not name Apache, a license, or 2.0.
The Preview weights are not public on Hugging Face. A footer link on the announcement pages goes to the Hugging Face organization page for Baidu, huggingface.co/baidu, which is an organization page rather than a Preview repository. The weights have not been released as open source yet.
Official demonstrations, and the text-only line
The official pages group demonstrations under game development and coding, 3D modeling and design, knowledge work, and video creation. Official copy says the model completed each piece end to end inside an agent harness, with external assets and tools credited alongside the work.
In game development, the announcement says the model has been adapted to major engines including Unreal Engine, Unity, and Godot, and highlights an 11-hour Unreal Engine task that produced Yunxiu Xing (《云岫行》), a third-person wuxia game set in Unreal Engine 5.8. In a single long coding session, the write-up says, the model merged assets made for several generations of Unreal Engine into one UE 5.8 project and wrote the game in C++. The same official account of that demo includes this line: "Because it is a text-only model and can't see the screen," the model turned playtest problems — such as a character hovering above the ground, an inconsistent running direction, and a misplaced camera — into measurable tests in code. The Chinese announcement uses the matching sentence 「作为纯文本模型，它看不到画面」. The tool-using demonstrations sit alongside that official description of a text-only model; one does not cancel the other.
The announcement also describes nearly 100 mobile mini-games built from complete design documents in Cocos, Unity, LayaAir, PixiJS, and Canvas 2D, with a video showing 24 of them. Other official examples include Blender work that built an airport and an airliner, and a model of Hōryū-ji; editable CAD assemblies in FreeCAD, including an engine, a model of Baidu International Building (Shenzhen), and a lunar-rover concept; a 20-page research report on AI data-center electricity costs and a 24-page report on autonomous driving; an interactive personal wealth-planning workbook; an original orchestral piece of about 6 minutes 34 seconds, delivered in WAV; and video work that includes a 116-second 2D short written in Python and a 75-second physics explainer produced from a single prompt. Official knowledge-work text mentions training environments built around large collections of reference files; official video-creation text mentions a pipeline from concept and script through visuals and voice-over. These are official demonstrations and official self-description published with the Preview, not independent evaluations.
Official benchmark charts
Baidu published two official PNG images with the announcement: a results chart and a full results table.
Baidu's official chart, titled "ERNIE 5.5 Flash Preview Benchmark Results." The note on the chart reads: "* indicates self-evaluated scores." These are Baidu's published official results, not independent tests.
Baidu's official full results table. Notes on the table state: "Metrics without any annotation are quoted from the model's official report or a public leaderboard; * indicates results from our own evaluation, and † indicates the text-only subset of HLE. All models are evaluated at their highest available reasoning setting." The comparison columns are Hy4, DeepSeek-V4.1-Flash-1017, GLM-5.3-Flash, GLM-5.3, GPT-6.1 Sol, and Claude Opus 5.5. On OfficeQA Pro V2, the official table lists ERNIE 5.5 Flash Preview at 67.9 and GPT-6.1 Sol at 70.0*. The table does not place ERNIE ahead on every row. These figures come from Baidu's official published table, not from independent tests.