On November 11, Baidu released ERNIE 5.5 Flash Preview, the first post-trained version built on the previously announced ERNIE 5.5 Flash Base. The model has 280B total parameters, activates 14B per inference and supports a 1-million-token context. The API is live on Baidu Qianfan, the official page also lists an OpenRouter entry, and the weights are due to be open-sourced by November 25.
Baidu positions the release as an intermediate checkpoint in post-training, and its release page says the model "may still show capability gaps caused by insufficient training." Post-training is still under way, and no timeline has been given for the final version.
The showcase centers on long-horizon agent tasks
The release page does not spend its space on benchmark scores. Evaluation results sit in an appendix, while the main body presents four kinds of work completed end to end inside agent frameworks.
Game development is the centerpiece. Baidu says it adapted the model specifically for Unreal Engine, Unity and Godot. In one showcase case, during an 11-hour long-horizon coding task, the model merged assets from several generations of Unreal Engine into a single UE 5.8 project and wrote a third-person wuxia game in C++. Because it is a text-only model and cannot see the screen, it turned playtest problems such as characters floating in the air, inconsistent running directions and misaligned cameras into measurable code tests, then iterated by adding tests round after round.
Another set of cases covers nearly a hundred mobile mini-games built on Cocos, Unity, LayaAir, PixiJS and others, spanning bullet-hell shooters, tower defense, card games and more; the video shows 24 of them. Baidu says most of the games separate the rule layer from the presentation layer, and that bot playthroughs and level replays were used to check automatically whether each game can be completed.
In 3D modeling, the model used the Blender Python API and FreeCAD scripts to build models including an airport with aircraft, Horyu-ji temple, a W16 engine and a lunar concept vehicle.
For knowledge work, Baidu gives the scale of its training environment as 300,000 workspaces, 10 million reference files and more than 40 file types. For video creation, it built more than 150,000 creative environments, with roughly 5,300 minutes of finished video produced in total.
All of these works were selected by Baidu, and the page notes the external assets and tools used in each, for example some visuals generated with Seedance 2.5 and some artwork coming from image models such as FLUX.2. How the model performs on everyday tasks will only become clear once developers get hands-on with it.
Three practical implications for developers in China
The first is the deployment barrier. With 280B total parameters, the weight files are large and the GPU memory requirement remains high. But only 14B parameters are activated per inference, so the compute at inference time is roughly comparable to a mid-sized dense model. MiniMax M3.1 Flash, which we covered earlier, activates 23B; ERNIE 5.5 Flash activates fewer, and what Baidu repeatedly stresses on the release page is "balancing intelligence and inference efficiency" along with decoding speed.
The second is the access route. The API runs on Qianfan, so Chinese enterprises can call it directly with an existing Baidu AI Cloud account, without dealing with the cross-border compliance issues of overseas models. Once the weights are open-sourced, they can also be pulled into a company's own data center. On the demand side, the second route is likely more attractive to financial and government customers with hard requirements that data stay in-domain.
The third is the license. Baidu's earlier ERNIE 4.5 series was open-sourced under the Apache 2.0 license, which allows commercial use. The release page does not say which license ERNIE 5.5 Flash will use or whether commercial use comes with extra conditions, so this cannot be confirmed for now.
What is still missing
Baidu does not provide an item-by-item comparison with other Chinese open-source models in the main text; benchmark scores are gathered in the appendix, and cross-model rankings will have to wait for third-party evaluations. The official side has also not said how large the gap between Preview and the final version will be, or when the final version will ship.
Pricing is a blank as well. The release page only gives entry points for Qianfan and OpenRouter, with no per-million-token prices. Whether the 1-million-token context can be used to the full reliably in real calls, and whether very long inputs will be billed separately, also depends on developer testing and official billing documentation. For teams planning to replace an existing model with this one, local inference performance after the weights are released may be a more useful reference than this round of showcase videos.
Sources: Baidu ERNIE official release page, CocoLoop, ERNIE 4.5 open-source notes; parameter scale, number of training environments and open-source deadline are as stated by Baidu.