ERNIE 5.5 Flash Preview: API, Free Access and Benchmarks

Last updated: November 11, 2026

Baidu released ERNIE 5.5 Flash Preview on November 11, 2026. It is the first post-trained model built on ERNIE 5.5 Flash Base: 280B total parameters, 14B active and a 1M-token context, tuned for coding, 3D modeling, knowledge work and content creation. The API is live on Qianfan, a free variant is on OpenRouter, and open weights are promised.

What’s new in ERNIE 5.5 Flash Preview
Flash Base vs. Flash Preview
How to call it on OpenRouter
Baidu’s benchmark results
Power-user workflows
Limitations to know
Frequently asked questions

What’s new in ERNIE 5.5 Flash Preview

Last week Baidu shipped a raw pre-trained checkpoint, covered in our ERNIE 5.5 Flash Base guide. The Preview is the first post-trained, instruction-following model on top of it, so it is the one you can actually chat with and run in agents (Baidu release page).

Baidu highlights four areas of targeted training:

Game development: domain-specific data and an evaluation framework for game mechanics, gameplay logic, runtime state and asset pipelines, with stronger support for Unreal Engine, Unity and Godot.
3D modeling and design: specialized training on Blender and CAD tools for parametric modeling and spatial reasoning.
Knowledge work: 300,000 training workspaces holding 10 million reference files in more than 40 file types, aimed at finance and legal deliverables such as spreadsheets, documents and slide decks.
Video creation: more than 150,000 creation environments covering script, storyboard, assets and voice-over.

Baidu calls the release an intermediate post-training checkpoint. It says some capabilities may still fall short and that it will improve the model for the official ERNIE 5.5 Flash release.

Flash Base vs. Flash Preview

Item	ERNIE 5.5 Flash Base (Nov 5)	ERNIE 5.5 Flash Preview (Nov 11)
Checkpoint	End of pre-training, no post-training	First post-trained version (intermediate checkpoint)
Size	280B total / 14B active	280B total / 14B active
Context	1M for fine-tuning and inference	1M tokens (OpenRouter lists 1,024,000)
Best use	Fine-tuning and research	Coding, agents, 3D, knowledge work, video
Access	Weights on Hugging Face and ModelScope	Qianfan API; free OpenRouter variant
Weights	Released under Baidu’s model license	Apache 2.0 promised; Chinese page says before November 25

Base details come from our earlier guide. Preview details come from the English and Chinese release pages and the OpenRouter model page.

How to call it on OpenRouter

OpenRouter lists one endpoint, served by Baidu, as baidu/ernie-5.5-flash-preview:free. It shows a price of $0 for input and output, a 1,024,000-token context, up to 921,600 completion tokens, text in and text out, and support for tools, tool_choice, structured outputs, response_format and reasoning controls (OpenRouter).

Create an OpenRouter account and generate an API key.
Point any OpenAI-compatible SDK at the base URL https://openrouter.ai/api/v1.
Set model to baidu/ernie-5.5-flash-preview:free.
Optionally add "reasoning": {"effort": "high"}. OpenRouter accepts effort values from none and minimal up to xhigh and max, and reasoning tokens count against your output limit (OpenRouter reasoning docs).
For agents, pass a tools array. The endpoint supports tool_choice values auto, required and a named function, but not none.
Handle HTTP 429. Free variants have per-minute and per-day request caps that rise once your account has purchased credits (OpenRouter limits).

You can also use Baidu’s Qianfan console instead. Baidu links the model card there, but it did not publish Qianfan pricing on the release page. For routing tips across providers, see our OpenRouter Auto Code guide.

Baidu’s benchmark results

Baidu published a chart comparing the Preview with larger and peer models. Baidu’s chart marks many scores with an asterisk as self-evaluated, so treat all of them as vendor claims until independent tests appear.

Benchmark	ERNIE 5.5 Flash Preview	Claude Opus 5.5	GPT-6.1 Sol	DeepSeek V4.1 Flash 1017
Terminal-Bench 4.0	46.5	63.6*	56.1	40.9*
DeepSWE v1.1	71.2	74.2	75.2	77.0
AutomationBench	61.9	58.0*	49.3*	57.2*
SpreadsheetBench V2	36.4	35.1*	32.0*	34.7*
OfficeQA Pro V2	67.9	62.2	70.0*	63.4*
Agents’ Last Exam (CLI)	29.0	32.4*	29.5*	35.2*

The pattern is clear: the Preview leads its chart on automation and spreadsheet tasks, but trails on terminal coding and long agent exams. Compare it with the models in our DeepSeek V4.1 Flash and Tencent Hy4 guides before choosing a default.

Power-user workflows

Turn visual bugs into tests. Because the model cannot see the screen, Baidu’s 11-hour Unreal Engine 5.8 demo converted playtest issues, such as a character hovering above the ground, into measurable code tests. Use the same pattern: describe the defect, ask for a failing test, then ask for the fix.

Script Blender and FreeCAD. Baidu’s demos used the Blender Python API and FreeCAD Python scripts to build editable scenes and assemblies. Ask for parametric scripts with named dimensions so you can change them later.

Use deterministic game rules. In Baidu’s mini-game demo, most games separated rules from presentation, so bots could replay levels and confirm each one can be cleared.

Split long jobs across agents. The airport demo ran two agents, one per subproject, then matched scale and cameras at the end. With 1M tokens of context, you can keep a full spec and file tree in each agent’s prompt.

For a domain-tuned alternative from another Chinese lab, see our GLM-5.3-FlashX-Legal guide.

Limitations to know

Text only. Image input requires an external tool, as in Baidu’s CAD demo that used an MCP visual-understanding tool.
Training is not finished. Baidu expects behavior to change before the official release.
Weights are not out yet. Apache 2.0 weights are promised, with a Chinese-page target of before November 25.
No published Qianfan price on the release page, and OpenRouter shows no paid endpoint yet.
Many scores in Baidu’s benchmark chart are marked as self-evaluated.

Frequently asked questions

What is ERNIE 5.5 Flash Preview?

It is Baidu’s first post-trained model built on ERNIE 5.5 Flash Base, released November 11, 2026. It has 280B total parameters with 14B active, a 1M-token context, and targets coding, 3D modeling, knowledge work and content creation.

Is ERNIE 5.5 Flash Preview free to use?

OpenRouter lists a free variant, baidu/ernie-5.5-flash-preview:free, at $0 per input and output token. Free variants on OpenRouter are subject to per-minute and per-day request limits that depend on how many credits your account has purchased.

How do I call ERNIE 5.5 Flash Preview through an API?

Baidu says the API is available on its Qianfan platform. On OpenRouter, send a standard chat request to https://openrouter.ai/api/v1 with the model ID baidu/ernie-5.5-flash-preview:free and your OpenRouter API key.

Will ERNIE 5.5 Flash Preview be open source?

Baidu’s English page says the weights will be open-sourced under Apache 2.0 soon. The Chinese page gives a target of before November 25. No weights were published at the time of writing.

Can ERNIE 5.5 Flash Preview read images?

No. Baidu describes it as a text-only model, and OpenRouter lists text input and text output only. In Baidu’s demos, visual checks were handled through tests in code or an external visual-understanding tool.

ERNIE 5.5 Flash Preview: Power-User Guide to API, Free Access and Benchmarks

Contents