Zhi东西（WeChat official account: zhidxcom）
Author | Cheng Qian
Editor | Xin Yuan
Zhi东西 11月11日消息，今日，Baidu released its first post-training version of the base model ERNIE-5.5-Flash-Base, named ERNIE 5.5 Flash Preview. The model has a total of 280B parameters and 14B activation parameters, supports million-context, and generates at a speed exceeding 200 tokens per second.
According to the benchmark test results released by Baidu, ERNIE 5.5 Flash Preview outperforms DeepSeek-V4.1-Flash-1017 in 14 benchmarks.
The model API has been launched on the Qianfan platform, and the weights will be open-sourced before November 25.
In terms of pricing, ERNIE 5.5 Flash Preview uses time-based pricing. From today to November 30, there is a 50% discount. During peak hours, the input cost is 1 yuan per million tokens, 0.1 yuan per million tokens hit in cache, and output cost is 4 yuan. During off-peak hours, the input cost is 0.5 yuan per million tokens, 0.05 yuan per million tokens hit in cache, and output cost is 2 yuan.
The original price was 2 yuan per million tokens input during peak hours, 0.2 yuan per million tokens hit in cache, and 8 yuan for output. During off-peak hours, the input cost was 1 yuan per million tokens, 0.1 yuan per million tokens hit in cache, and 4 yuan for output.
▲Price comparison of ERNIE-5.5-Flash-Preview and DeepSeek-V4.1-Flash-1017 (Zhi东西 illustration )
I. Without relying on image input, relying on code self-testing iteration, Baidu's new model can be tested and tuned
Baidu officially released several cases where ERNIE 5.5 Flash Preview completed end-to-end tasks in the Agent framework.
Firstly, in the field of game development, Baidu stated that ERNIE 5.5 Flash Preview can cover a wide range of programming scenarios, from fixing and iterating through real codebases to fully building complex applications.
In game development, researchers have enhanced the model's capabilities in areas such as implementing game mechanics, arranging core logic, managing runtime states, and handling resource pipelines by constructing high-quality specialized training data and corresponding evaluation systems. They have also adapted the model for mainstream game engines such as Unreal Engine, Unity, and Godot.
The following is an Eastern fantasy game called
《云岫行》created using the Unreal Engine.
During an 11-hour long-range coding task, ERNIE 5.5 Flash Preview integrated materials spanning multiple versions of the Unreal Engine into a single UE 5.8 project and wrote the entire game in C++.
Baidu also showcased cases of the model generating various mobile games.
ERNIE 5.5 Flash Preview generates games such as bullet-screen shooting, roguelike, tower defense, racing, card RPG, city-building and management, physics-based puzzles, and interactive storytelling based on different game engines. In constructing games, the model separates the rule layer and presentation layer, creating a deterministic simulation, then allows robots to playtest, perform state space search, and replay levels to automatically check if levels can be cleared and if numerical balance is achieved.In terms of 3D modeling, researchers enhanced specialized training on tools such as Blender and CAD, enabling the model to understand and infer the shape, proportions, and spatial relationships of objects from text descriptions, converting them into 3D models through procedural modeling, while also supporting integration and reuse of existing assets.
ERNIE 5.5 Flash Preview drives two Agents to create an airport and an airplane separately using the Blender Python API.
For the airport part, the model retrieves layout vector data from the map platform, downloads auxiliary materials such as plants, vehicles, and ground service equipment as needed, and uses parametric modeling to build the terminal and flight area. For the airplane part, the model creates the fuselage and wings using section extrusion and modular assembly, then assembles the cockpit, double-deck cabin, and cargo hold using pre-provided generative materials. Subsequently, the model unifies the scale, position, and camera angles of both projects, arranges the presentation animation, and renders it frame by frame using Cycles/OptiX GPU, then composites the video using FFmpeg.
Another case involves using CAD to design an engine model. ERNIE 5.5 Flash Preview created an editable engine model in FreeCAD using Python scripts, including main components such as the cylinder block, crankshaft, piston, and connecting rod.
▲ When the crankshaft rotates, the connecting rod drives the pistons in the four cylinder columns to reciprocate.
In the domain of knowledge work, ERNIE 5.5 Flash Preview can independently complete multi-step tasks to deliver tables, documents, or presentations. Focused on fields such as finance and law, Baidu has built over 300,000 workspaces and 10 million reference files, covering more than 40 file types, with training environments ranging from personal workspaces with 10 to 100 files to enterprise knowledge bases consisting of thousands of files.
In the field of video creation, Baidu has built over 150,000 creation environments for scenarios such as science popularization, manga, creative editing, and hot topic performance, covering the complete process from creative generation, scriptwriting, storyboard, visual materials, to voiceover, adapting to thousands of creation skills, and accumulating approximately 5,300 minutes of completed videos.
Below is an example of ERNIE 5.5 Flash Preview writing a 116-second 2D animated short film using Python. A white pelican rides a bicycle, traveling from the rock paintings of the Sahara in 4000 BC all the way into the code editor of the large model era, crossing through 20 art styles, and finally arriving at the "21st Pelican Release Event".
Baidu revealed that some of the background scenes in this case were completed by ERNIE 5.5 Flash Preview calling generative image tools, including static base maps for the Tang Dynasty sketch, Qingming Shanghe Tu, and Mughal miniature painting scenes, all of which originated from the generative image tool, with ERNIE 5.5 Flash Preview then overlaying animated characters, falling flowers, and water ripples on the base maps using code.
Second, it outperformed DeepSeek-V4.1-Flash-1017 in 14 tests, achieving the top rank in multi-tool call evaluation.
According to Baidu's publicly disclosed benchmark test results, it compared ERNIE 5.5 Flash Preview with Hy4, DeepSeek-V4.1-Flash, GLM-5.3-Flash, GLM-5.3, GPT-6.1 Sol, and Claude Opus 5.5.
ERNIE 5.5 Flash Preview took first place in the MCP protocol multi-tool call evaluation MCP-Atlas, which tests the model's ability to connect various external tools and orchestrate toolchains. In the web browsing Agent evaluation BrowseComp and the DRACO long process business Agent evaluation, the Baidu new model scored second, just behind Hy4.
The model performed better than DeepSeek-V4.1-Flash-1017 in 14 out of the following 20 benchmark tests, including tool call tests MCP-Atlas, BrowseComp, browser Agent, office automation, CookieBench, and pure reasoning HLE.
Conclusion: Goodbye to one-model-for-all, large models enter the era of layered scheduling
Recently, Baidu's model iteration has accelerated. On November 5th, they first open-sourced the Base foundation model, then released the first post-training preview version within a week, and on November 25th, they open-sourced the training weights. Now, the division of labor among large models is becoming increasingly clear, with flagship models handling more difficult tasks, and Flash models focusing on processing high-frequency, repetitive tasks. This Flash preview version released by Baidu demonstrates greater cost-effectiveness. From the company's layout, it can be seen that they are forming a combination of models including the foundation model, flagship model with strong reasoning, and Flash model for fast reasoning, to meet different user needs.
However, Baidu disclosed that ERNIE 5.5 Flash Preview is an intermediate version in its post-training process, and the post-training is still ongoing. ERNIE 5.5 Flash is still rapidly evolving. Although it has already achieved relatively competitive performance among models of the same size, there may still be some capability deficiencies caused by insufficient training.
Source: Baidu
