Baidu released ERNIE 5.5 Flash Preview on November 11, 2026, introducing a large but computationally selective language model designed for coding, agentic workflows, document-heavy knowledge work and software-assisted creation. Its headline specifications are 280 billion total parameters, 14 billion active parameters per token and a one-million-token context window. The model is accessible through Baidu's Qianfan API, while the weights of this preview were not publicly downloadable at launch.
The announcement is significant for developers working with long documents and multi-step tools, but it should not be mistaken for a change to Baidu Search. Baidu has not demonstrated that this preview changes which websites its public search engine cites or how much referral traffic publishers receive.
What Baidu has released
According to Baidu's official ERNIE announcement, ERNIE 5.5 Flash Preview is the first post-trained model built on the ERNIE 5.5 Flash Base checkpoint released earlier in November. Baidu describes the preview as an intermediate post-training version rather than a finalized model. Its examples focus on game development, 3D design, document production, research and tool-using agents.
Baidu's official Qianfan product page lists the model as text-to-text, with 280B total parameters, 14B activated per token, a one-million-token context and claimed generation speeds exceeding 200 tokens per second. That throughput is a vendor claim tied to its serving environment, not a guarantee across every prompt, region or third-party provider.
ERNIE 5.5 Flash Preview specifications
| Attribute | Documented detail |
|---|---|
| Developer | Baidu |
| Release date | November 11, 2026 |
| Model type | Text-in, text-out language model |
| Architecture | Mixture-of-Experts, with selective parameter activation |
| Total parameters | 280 billion |
| Active parameters | 14 billion per token |
| Context window | Approximately one million tokens |
| API identifier | ernie-5.5-flash-preview |
| Primary access | Baidu Qianfan API |
| Additional access | OpenRouter lists a hosted version, subject to its own terms |
| Weights | Preview weights not publicly released at launch |
| Planned license | Apache 2.0 for a future weights release, according to Baidu |
| Final release date | Not confirmed in the launch documentation |
Parameter counts do not directly measure model quality. A mixture-of-experts architecture can maintain a large overall capacity while activating a smaller portion of its parameters for each token. Actual latency, memory requirements and performance also depend on routing, attention design, hardware, batching and serving software.
Why a one-million-token context matters
A one-million-token window makes it possible to submit very large document sets, code repositories, support histories or research materials to one model invocation. For example, an organization might ask an agent to compare policy changes across hundreds of documents, identify conflicting contract terms or prepare a spreadsheet from extensive reference material.
However, a large context window does not prove reliable long-context reasoning. Models can miss relevant passages, confuse similarly named entities or produce unsupported summaries. Teams should measure answer accuracy, citation fidelity and retrieval from different positions in the input rather than assuming that every supplied token is used correctly.
Agentic coding, 3D design and knowledge work
Baidu highlights workflows in which ERNIE 5.5 Flash Preview writes code, operates development tools and iterates on outputs. Its demonstrations include game creation with Unreal Engine, automated modeling with Blender or CAD software, and producing documents, slides and spreadsheets using external applications.
These are examples of a model operating inside a tool-equipped agent environment. They are not evidence that the base text model can directly see a computer screen or generate video pixels. Baidu's release material identifies the preview as text-only, while its showcased creative workflows use external tools and assets. This distinction is especially important when comparing it with genuinely multimodal input or video-generation models.
The company also describes post-training on large collections of executable coding tasks and simulated professional workspaces, including finance and legal materials. Such training may improve tool-use behavior, but real deployment still requires permission controls, error handling, reproducible tests and human review for consequential outputs.
Benchmarks: promising claims, limited independent validation
Baidu has published benchmark comparisons for coding, agent tasks and professional work, including results on evaluations such as Terminal-Bench, MCP-Atlas and BrowseComp. These results are primarily vendor-supplied, and some comparison scores may have been reproduced under Baidu's own settings. They should not be presented as an independent consensus ranking.
Third-party model listings and emerging arena evaluations may provide additional signals, but small or early samples can be volatile. A reliable comparison requires consistent reasoning budgets, identical tools, comparable latency and cost measurements, and tests on actual target workloads. At launch, no broad independent evaluation established that ERNIE 5.5 Flash Preview was superior across every task.
API availability, pricing and the promised open weights
The model can be called through Baidu's Qianfan platform using ernie-5.5-flash-preview. Developers should consult the current model list for account eligibility, regional endpoints, rate limits and the exact supported request parameters.
Baidu's Qianfan marketing page advertises introductory input pricing from RMB 0.5 per million tokens and output pricing from RMB 2 per million tokens under promotional conditions. These are not universal prices for every provider or time of day. Third-party hosted listings, including OpenRouter, may have different USD-denominated rates and caching policies. Before calculating production cost, verify current billing, context-dependent rates, cache treatment and tool-related charges.
Baidu says the model weights will be released under the Apache 2.0 license. The Chinese-language announcement and subsequent official social posts point to a target before or by November 25, but the key editorial distinction remains: the preview API is available now; downloadable preview weights were not yet available at launch. A promised open-weight release is not the same as an existing model repository. Developers should verify the actual downloadable files, license and deployment requirements when Baidu publishes them.
How to test the model through Qianfan
- Open the official Baidu Qianfan model catalog and confirm that
ernie-5.5-flash-previewis available to your account and region. - Review current API authentication, pricing and rate-limit documentation before creating a test key.
- Run a small text-generation request to validate the endpoint and model identifier.
- For tool use, define a narrow, allowlisted set of functions and inspect every requested action.
- Test long-context retrieval using documents with known answers and verifiable source locations.
- Measure end-to-end latency, generated tokens, tool failures and cost per successfully completed task.
- Keep the model identifier configurable because a preview may change or be superseded.
Do not assume that a demonstration showing an agent using Blender, spreadsheets or a browser means those tools are automatically available in a basic Qianfan text API call. The surrounding application must provide the tools and permissions.
What this means for SEO and AI Search
The release has indirect relevance to GEO and AI visibility. A model with a large context window and strong tool-use capabilities can help developers build research assistants, enterprise search tools and agents that process public webpages alongside private documents. In such applications, source selection and attribution may influence which publishers or brands are represented in the final answer.
But this is not a confirmed Baidu Search ranking update. The announcement does not say that Baidu Search has replaced its public search-answer model with ERNIE 5.5 Flash Preview, and it supplies no before-and-after dataset on citations, impressions, clicks or referral traffic. A model release, an API integration and a search-product deployment are three different events.
Why source attribution matters in agentic workflows
Consider a research agent comparing industrial suppliers. It may receive thousands of pages of specifications, use tools to query updated catalogs and generate a structured shortlist. The answer's usefulness depends on whether the agent can distinguish current product information from outdated marketing claims and link each recommendation to supporting evidence.
That creates an opportunity for websites that publish precise, dated, attributable and actionable information. Yet it does not justify claiming that a particular metadata tag or writing style guarantees inclusion in ERNIE-powered answers. Source selection is controlled by the actual retrieval and agent implementation, which may differ between Qianfan applications and Baidu's consumer products.
Practical GEO tests for researchers and publishers
A repeatable evaluation can establish whether a specific ERNIE-powered application retrieves or uses a website:
- Select 20–30 queries covering brand identity, product facts, comparisons and transactional tasks.
- Record whether the application has web access, an enterprise document index or no retrieval tools at all.
- Use identical prompts and comparable source corpora across tested models.
- Log retrieved URLs separately from URLs cited in final answers.
- Measure factual accuracy, attribution, source prominence, task completion and latency independently.
- Include multilingual queries where relevant, and distinguish Chinese-language from English-language performance.
- Test whether the agent ignores instructions embedded in untrusted documents or webpages.
- Track qualified website visits and conversions separately from mere AI mentions.
These measurements can establish application-specific visibility. They cannot, by themselves, establish a change in Baidu's public search algorithm.
What remains unconfirmed
| Question | Status |
|---|---|
| Did Baidu release ERNIE 5.5 Flash Preview on November 11? | Confirmed. |
| Does it have 280B total and 14B active parameters? | Officially documented. |
| Does it support a one-million-token context? | Officially documented. |
| Can developers access it via Qianfan? | Yes, subject to account and regional conditions. |
| Are the preview weights already downloadable? | Not at launch; a future Apache 2.0 release is promised. |
| Has Baidu announced a final general-release date? | Not established by the launch documentation. |
| Has it changed Baidu Search citations? | No demonstrated change. |
| Has it increased publishers' organic clicks? | No supporting dataset provided. |
Frequently asked questions
What is ERNIE 5.5 Flash Preview?
It is Baidu's November 2026 preview of a post-trained mixture-of-experts language model designed for coding, knowledge work and agentic applications.
How many parameters does it have?
280 billion in total, with 14 billion active per token, according to Baidu.
What is its context window?
Approximately one million tokens.
Is ERNIE 5.5 Flash Preview multimodal?
The model is described as text-to-text. Demonstrations involving video, games and 3D design rely on external software tools and do not prove native image or video understanding.
Where can developers access it?
Through Baidu's Qianfan API, with additional third-party hosting listings subject to their own terms.
Is ERNIE 5.5 Flash open source?
The preview weights were not public at launch. Baidu says an Apache 2.0 weights release is planned.
When will the weights be released?
Baidu's Chinese announcement and subsequent official communications point to November 25, 2026, as a target. Actual availability should be verified from the downloadable repository.
Can it call tools?
Baidu showcases agentic workflows and tool use. Specific tool capabilities depend on the API and surrounding agent framework.
Is it the default model for Baidu Search?
No such deployment was established by the preview announcement.
Will it change SEO rankings?
There is no evidence of a direct ranking change. Its relevance is potential use in retrieval and agentic applications.
The bottom line
ERNIE 5.5 Flash Preview adds another substantial model to China's fast-moving AI ecosystem, combining a 280B-parameter MoE architecture, one-million-token context and agent-oriented post-training. The Qianfan API is available, and Baidu has promised open weights under Apache 2.0. For SEO and GEO teams, the responsible conclusion is narrower: this is a model and developer-platform release, not a proven change to Baidu Search citations or publisher traffic. Its real impact will depend on independent testing and the products that eventually deploy it.
Sources
- Baidu ERNIE — Introducing ERNIE 5.5 Flash Preview, November 11, 2026 (official announcement).
- Baidu AI Cloud — ERNIE 5.5 Flash Preview on Qianfan (official model specifications and promotional pricing).
- Pandaily — Baidu releases ERNIE 5.5 Flash Preview, November 11, 2026.
- DigIntoAI — ERNIE 5.5 Flash Preview release and Chinese-language announcement details, November 11, 2026.