Baidu released ERNIE 5.5 Flash Preview on November 11, 2026. It is a test version of the company's newest large language model, and developers can call it through an API today. Baidu says the open weights will follow under the Apache 2.0 license.
The release matters because it turns the raw base model Baidu shared earlier into a model that people can actually chat with and give tasks to. It also comes with API prices and a date for the weights, so readers can check whether Baidu keeps its word.
What did Baidu release?
On its ERNIE release page, Baidu calls ERNIE 5.5 Flash Preview "the first post-trained model built on the ERNIE 5.5 Flash Base we released earlier." We covered that base model when its weights came out: Baidu Opens ERNIE 5.5 Flash Base.
"Post-trained" means Baidu took the base model, which only predicts text, and trained it further to follow instructions, use tools and finish multi-step jobs.
Baidu gives these specs:
- 280B total parameters, with 14B active for each token (a mixture-of-experts design, where only part of the model runs at a time)
- A context window of 1 million tokens
- Strong areas, in Baidu's words: coding, 3D modeling, knowledge work and content creation
Xinhua adds more detail. It says the model uses a new hybrid sparse attention design, a multi-router MoE and ECHO multi-token prediction, and that it generates more than 200 tokens per second. Xinhua also says the model uses Baidu's in-house "RSI" technique, with ERNIE Evolve taking part in inference optimization. ERNIE Evolve is the agent system we explained in Baidu ERNIE Evolve: AI Agents Speed Up ERNIE 5.5 Flash.
What does Baidu say it can do?
Most of the release page is a set of demos. Baidu says the model completed each one end to end inside an agent harness. These are Baidu's own showcase results, not independent tests.
- Games. Baidu says the model wrote a third-person wuxia game in C++ for Unreal Engine 5.8 in "a single 11-hour coding task." Because it is a text-only model and cannot see the screen, it turned playtest problems, such as a character hovering above the ground, into measurable tests in code. Baidu also says it built nearly 100 mobile mini-games from game design documents.
- 3D and CAD. Two agents running the model used the Blender Python API to build an airport and an airliner. Other demos show the Hōryū-ji temple in Blender, plus a W16 engine and a lunar rover concept modeled with Python scripts in FreeCAD.
- Office work. Demos include a 20-page research report on AI data center electricity costs and an Excel wealth planner that runs a Monte Carlo simulation over 1,000 randomized market paths.
- Video. The model wrote a 116-second animated short in Python, which Baidu calls "The Pelican Test."
Baidu also describes how it trained these skills. For domains such as finance and law, it says it built 300,000 workspaces that hold 10 million reference files in more than 40 file types. For video, it says it built more than 150,000 creation environments.
How much does it cost and how can you try it?
The API is live on Baidu's Qianfan cloud platform. Xinhua gives these prices per million tokens:
- Input: 2 yuan at peak times, 1 yuan off-peak
- Output: 8 yuan at peak times, 4 yuan off-peak
- A limited-time 50 percent discount on top
Baidu's release page also links to OpenRouter. On November 11, 2026, OpenRouter's model list showed a free variant, "baidu/ernie-5.5-flash-preview:free," priced at $0 with a 1,024,000-token context. Free variants on OpenRouter can change or end, so check the listing before you build on it.
When will the weights be released?
Baidu's English release page says only that the weights "will be open-sourced under Apache 2.0 soon." Its Chinese release page says the weights will be open-sourced by November 25, and Xinhua says they will go to the global open-source community on November 25, 2026. That is 14 days after the preview launch. We will check on that date and add an update.
If the weights ship under Apache 2.0, anyone could download, run and change the model on their own hardware under that license's terms. For the wider ERNIE family, see our guide to ERNIE and Baidu's AI models.
What to watch
- It is not finished. Baidu calls the Preview "an intermediate checkpoint from post-training." It says that "because training is not finished, some capabilities may still fall short," and that it will make improvements in the official release. Baidu gave no date for that release.
- The claims are Baidu's. The demos come from Baidu, and the speed figure comes from Xinhua's report. Baidu's page includes a benchmark appendix, and Xinhua says the model scored competitively on Terminal-Bench and AutomationBench, but no outside group has checked those results yet.
- The weights date is a promise. Until files appear on Baidu's model pages, the open-weight part of this release is a plan, not a fact.
For now, ERNIE 5.5 Flash Preview is an API model with a 1 million token context window that Baidu pitches at agent work in code, 3D, documents and video. The November 25 weights release is the next thing to check.
