ERNIE 5.5 Flash Preview opens free API access, with weights still pending
Baidu has introduced ERNIE 5.5 Flash Preview, with API access available now and downloadable weights still to come. Developers can also try a free endpoint on OpenRouter, subject to its usage limits.
The dates need a little care. Baidu’s official site dates the announcement November 11, 2026, while OpenRouter’s listing gives November 10. Neither establishes a precise global launch time. This is a newly announced preview, rather than proof that every access route opened simultaneously.
Where you can use ERNIE 5.5 Flash Preview
| Access route | What is available | What to check |
|---|---|---|
| Qianfan | Baidu says the API is live | Account eligibility and current provider pricing |
| OpenRouter | Free API endpoint served by Baidu Qianfan | Request limits, capacity and routing requirements |
| Local deployment | Preview weights are forthcoming | No verified downloadable release yet |
The exact OpenRouter identifier is baidu/ernie-5.5-flash-preview:free. Its listing shows zero input and output token charges, a 1,024,000 token context window, function calling and structured output support. It also says this endpoint is unavailable through routing that stays within the EU. That routing restriction should not be generalized into a claim that every European user is blocked.
Free access is useful for an initial comparison, but budget for the whole experiment. A coding agent can make many requests while investigating files, running tests and repairing errors. Any external services that your own tools call have separate costs and permissions.
Free API access still has limits
OpenRouter’s rate limit documentation distinguishes platform request caps from capacity limits at the upstream provider. Free usage allowances depend on account policy. The documentation points to the authenticated key information endpoint for the remaining daily allowance and recommends honoring retry instructions when a request is limited.
Check those limits before starting an unattended job. Keep a record of the model identifier, provider, request settings and error response. If a long task stops, that record helps separate a capacity problem from a model failing to complete the assignment. Repeatedly submitting the same failed request is a poor recovery plan.
A text model coordinating creative tools
Baidu describes an intermediate checkpoint from post training with 280 billion total parameters, 14 billion active per token and a one million token context. Its launch examples include game development, Blender scenes and documents produced through an agent harness. These are vendor demonstrations, not ByteForward tests.
One game development example describes an 11 hour Unreal Engine task that combined existing assets with generated C++ code. Baidu explicitly says the model is text only. The surrounding tools and feedback matter when interpreting the resulting visuals.
That distinction changes how to evaluate it. Ask for the editable project, the dependency list and repeatable checks, rather than judging a polished video alone. For a small game, a useful trial would include restoring a saved state, changing a rule and checking that an existing level still works. Those are suggested acceptance checks, not tests we ran.
A large context window also needs a practical test. Try an existing repository with relevant information spread across several files, then check whether the model uses the right version of each fact. Capacity tells you how much material a service accepts. Your own task determines whether it uses that material reliably.
Keep Preview and Flash Base separate
Our ERNIE 5.5 Flash Base guide covers the earlier pretraining checkpoint and its custom license. For Preview, Baidu promises a future Apache 2.0 weights release. Do not apply that promise retroactively to the Base download, or treat hosted API access as a local deployment package.
For now, the sensible starting point is a contained API evaluation using disposable project copies and narrowly permitted tools. Record how much correction the task needs and whether the final artifact can be reopened and edited. Wait for an actual weights repository and its license before planning self hosting around this Preview release.
Featured image: AI generated editorial illustration by ByteForward, using the official Baidu and ERNIE marks. Conceptual artwork, not a product screenshot. No endorsement implied.
