Is Baidu Wenxin’s recent efforts in catching up strong enough?
Author | Zhou Lei, Lin Jiexin
Editor | Cen Feng
Today, Baidu Wenxin 5.5 Flash Preview was officially released, marking the third major update in just one week.
@@ On November 5th, the underlying model Wenxin 5.5 Flash Base was made open-source;
On November 9th, progress on the RSI (Research Speed Indicator) project was announced, along with the release of ERNIE Evolve;
On November 11th, Wenxin 5.5 Flash Preview was officially launched.
Three major updates in one week is quite unusual in Wenxin’s history. These actions are closely linked: from the development of the underlying model to the exploration of the RSI indicator, and finally to the release of the preview version, gradually revealing new directions for Baidu’s foundational model business.
In the public’s perception, Baidu Wenxin’s position has been somewhat ambiguous: the project was launched early and has made significant progress, but its presence in the market has not been as prominent as expected. In the competitive landscape of large-scale language models, Wenxin has not yet met the expectations set for it.
In July of this year, Sun Tianxiang, a Ph.D. from Fudan University born in 1997, joined Baidu and took charge of the Basic Model Unit (BMU), as well as the Baidu Model Committee (BMC). He was previously the lead developer of the domestically produced, open-source ChatGPT-like model MOSS and has also started his own business. The public is looking forward to his first report on the progress of Baidu’s foundational model efforts. This series of three updates this week may be part of that report.
The newly released Wenxin 5.5 Flash Preview focuses on efficiency, cost-effectiveness, and open-source principles. The model is now available on the Qianfan platform, with both the web interface and APIs being made publicly accessible. According to the official statement, the model’s weights will be released to the global open-source community under the Apache 2.0 license by November 25th.More information about the models will also be disclosed at Baidu World 2026 tomorrow. Let’s start by looking at the rankings. Wenxin 5.5 Flash Preview achieved first place in two of the eight benchmark tests: AutomationBench and SpreadsheetBench V2. Among the models included in the comparison, DeepSWE, APEX-Agents, DRACO, and OfficeQA Pro V2 ranked second in those four tests, while the other models also performed well. Baidu’s official description of these results is quite modest, using the phrase “achieved competitive scores.” It should be noted that the scores marked with an asterisk in the comparison chart are based on the manufacturers’ own evaluations and are for reference only. In addition to the ranking positions, the planning of the model’s capabilities and the engineering details during the testing process provide a better understanding of the model’s actual performance. According to Baidu’s official statement, Wenxin 5.5 Flash Preview has been upgraded in four main areas: programming, 3D modeling, knowledge work, and content creation. For programming, it was trained using over 14 million executable code tasks and various Agent frameworks; for 3D modeling, it was specifically trained for tools such as Blender and CAD; for knowledge work, it was tested in 300,000 workspaces across fields like finance and law, involving 10 million files; for content creation, it was adapted to thousands of skills and more than 150,000 creative environments were built, with a total output of approximately 5,300 minutes of content. Leifeng Network and AI Technology Review immediately tested Wenxin 5.5 Flash Preview’s game development capabilities. Your task is to create an original 2D rhythm game called “PUNCH! RETURN!” that is inspired by the gameplay concept of “Rhythm Heaven.” Your team consists of experienced game producers, game designers, 2D art directors, character animators, music designers, and game engineers. The goal of the game is to create a complete, interactive, and vibrant cartoon musical comedy.The entire scene must resemble a truly existing, absurd cartoon world: not only players and delivery robots, but also various minor characters that appear continuously, each with their own independent actions; reactions to the players’ success or failure; a constantly changing environment; unexpected disruptions; and ultimately, a climactic moment in which all characters participate together. The following is the performance of Wenxin 5.5 Flash Preview during this test: the game successfully completed core tasks such as the gradual introduction of characters, timing judgment, and the conclusion of the narrative. After multiple rounds of generation, the total time taken was approximately 7 minutes and 32 seconds, resulting in a version that is capable of running. This test to some extent demonstrated Wenxin 5.5 Flash Preview’s ability to handle highly complex requirements. It also assessed the model’s aesthetic sense, musical rhythm, visual composition, and animation production capabilities. On the coding side, it was essential to ensure that all events were triggered correctly. For example, if the character state system was not well-designed, the character might not be able to recover from a tense state after consecutive failures, which could have a chain reaction throughout the entire scene. A single case study alone cannot fully reflect the stability of the model’s output, but tasks that involve multiple systems and require comprehensive documentation are among the most challenging tests of programming skills. The official blog’s demonstration cases cover areas such as video composition, game development, and 3D modeling, with game development being the most significant part. A representative example of game development is “Yunxiu Xing” (Cloudy Peaks Journey): a third-person Eastern martial arts game set in the misty landscapes of Zhangjiajie’s rocky peaks and ancient temples, built using Unreal Engine 5.8. The entire project took 11 hours to complete; the model integrated materials from multiple versions of the Unreal Engine into a single UE5.8 project, and the game was fully implemented in C++. During the trial play, visual issues such as characters appearing to be floating in mid-air, inconsistent running directions, and misaligned camera positions were identified and transformed into measurable code testing problems. The project was iterated through continuous testing. “Converting visual problems into testable code” indicates that the model has established a self-verifying mechanism, which is closer to the actual development process than simply generating a runnable version of the software – and this is precisely the most difficult aspect to automate in long-term development projects. The official documentation also includes larger-scale cases: based on complete game planning, the model was used to create nearly a hundred mobile games covering various genres such as bullet shooting, tower defense, racing, card RPGs, city-building management, physics-based puzzles, and interactive narratives. Twenty-four of these games were developed using multiple technologies including Cocos, Unity, LayaAir, PixiJS, and native Canvas 2D. Most of these games used a separation of the rule-based logic and the visual representation layers for more accurate simulations, and the model automatically verified the solvability and numerical balance of the games through robot simulations, state space searches, and level replays – a method of engineering-based validation that is consistent with the approach used in “Yunxiu Xing”. Video composition involves aspects such as rhythm, timing, and consistency. “Pelican Test” is a 117-second 2D animation short film: Wenxin 5.5 Flash Preview was used to create this animation from scratch using Python. In the film, a white pelican rides a bicycle from Saharan rock paintings from 4000 BC all the way into the code editor of the modern era, traversing through 20 different artistic styles, and finally appearing at the “21st Pelican Launch Event.”The entire video is driven by tens of thousands of lines of code; each frame is vector-drawn using Pycairo, with subsequent processing done by NumPy and Pillow. The same parametric pelican appears in 20 different scenes, taking on various forms such as a Han dynasty stone inscription, a Javanese shadow puppet, the light from an oscilloscope, or an 8-bit pixel image. As the scene changes, the image transitions smoothly, as if on film, with the pelican consistently crossing the seams between the frames. In the scene “Starry Night,” the swirling starry sky moves in sync with the pace of the video. The soundtrack consists of 22 selected recordings, ranging from conga drums, guqin music, and gamelan to the Mario theme; a script written by the model aligns these sounds with the corresponding scene transitions. For some backgrounds, the ERNIE 5.5 Flash Preview tool was used to generate custom images: the static backgrounds for the Tang dynasty-style drawings, the “Riverside Scene at Qingming Festival,” and the Mughal-era paintings were all created using this tool, with the model then adding moving characters, falling flowers, and water patterns on top of them using code. The conceptual design was created by Bilibili creator @JinbuBaogao_. The quality of individual frames in video generation competitions has already reached a very high level; the real challenges lie in maintaining consistency across different scenes and controlling the overall rhythm of the video. The fact that the same pelican maintains its appearance across 20 different styles demonstrates the model’s ability to manage its state over a long period of time. In the 237-second official trailer, the model independently completed four tasks: researching resources, writing the script, designing the animation sequences, and generating the final video. According to official data, the model worked continuously for one to two hours to complete the video, with an average of about 170 tool calls during this process. During this time, the model also checked its own results, identified issues, and made necessary corrections. The 3D modeling for both cases was done using Blender. In one example, two intelligent agents based on the Wenxin 5.5 Flash Preview technology collaborated using the Blender Python API: one agent was responsible for the airport, independently retrieving layout data from map platforms, downloading authorized assets such as plants, vehicles, and ground services, and using parametric modeling to construct the terminal building and flight areas; the other agent was responsible for the airplane, building the fuselage and wings through cross-sectional drafting and modular assembly. Both agents then aligned their work in terms of scale, position, and camera perspective, animated the scenes, and used Cycles/OptiX for GPU-based rendering and FFmpeg for video synthesis. In the second example, the model independently researched the layout and architectural structure of the Falang Temple, gradually constructing the entire building complex, with attention to details such as the tapering of each floor, the support of beams and columns, and the arrangement of roof tiles. Licensed scans of Buddha statues and plant assets were integrated into the scenes, with the surface details completed using PBR textures and procedural materials. The common theme of both cases is that the modeling process was carried out within a complete “retrieve—model—render—compose” pipeline, rather than executing a single modeling command at a single point in time. The division of tasks and collaboration between the two intelligent agents also extended the capabilities of the model beyond what a single model could achieve on its own.
02
“快”……这速度是从哪里来的呢？
上面提到的游戏和视频案例都属于“长距离任务”：需要连续工作一到两个小时；在这些任务中，工具被调用的平均次数约为170次；整个项目迭代过程耗时11小时。这类任务对计算速度的要求远高于日常的问答场景，因此“快”在这里具有实际意义。
对于大多数用户来说，一个更直观的指标是：千帆API的单用户推理速度超过了200 tokens/秒，这意味着每秒钟可以处理超过150个汉字。当生成速度达到这个水平时，输出结果几乎可以实时跟上用户的阅读速度，从而显著减少用户的等待时间。
官方对“快”的解释是基于一套系统的技术设计，这套设计可以分为三个层次来理解。
第一个层次是参数的使用方式。文心5.5的Flash Preview功能总共使用了2800亿个参数，但其中只有140亿个参数在每次计算时被真正激活。在当前的配置下，这种设计已经相当激进了。总参数数量决定了模型的知识容量，而只有被激活的参数才会影响实际的计算成本。因此，2800亿个参数更像是一种“备用库存”：每次生成新的结果时，只需要调用其中的一小部分参数而已。
第二个层次是显存（RAM）的使用效率。如果对每个输入的“token”都使用全量的计算资源（即“全注意力”机制），显存的消耗将会非常高，导致系统无法正常运行。为了解决这个问题，采用了混合稀疏注意力（mixed sparse attention）机制：在43层计算结构中，有37层会将每128个输入token压缩成一个缓存条目，从而保留全局的上下文信息；剩下的6层则保留了对每个token进行单独访问的能力。这种设计大大降低了显存的消耗，同时仍然能够有效地检索长文本中的细节。
第三个层次是推理阶段的加速。FP4技术将专家权重（expert weights）的存储压缩到了4比特，从而节省了显存的带宽；ECHO技术则让主干网络的最后一层在预测下一个token的同时，也负责预测后续的token，这层网络同时还充当了一个“草稿模型”（a draft model for speculation），通过先进行猜测再进行验证的方式，几乎不增加额外的参数数量，就能显著提高计算速度。
这三层设计共同实现了同一个目标：使用更少的参数来提供足够的智能处理能力，从而提高计算速度并进一步降低成本。这也是“Flash”这一名称的由来——它强调的不是单一技术的极限性能，而是产品线在效率、能力和成本之间的平衡。
Baidu has released the complete technical blog and model architecture diagram during this synchronization process.
03
The pricing is open-source, and there is also a special signal involved.
In terms of pricing: the API costs 2 yuan per million tokens during peak usage and 1 yuan during off-peak hours for input; it costs 8 yuan for output during peak usage and 4 yuan during off-peak hours. Currently, there is a 50% discount, so the cost for input during off-peak hours is 0.5 yuan, and the cost for output is 2 yuan. This is a competitive price in models of the same parameter scale.
To put it more clearly: one million tokens is approximately equivalent to 600,000 to 800,000 Chinese characters. Taking the longest context supported by the model as an example, filling a 1-million-token window completely costs less than 1 yuan after the discount.
The efficiency has been designed to provide a low-cost alternative to Wenxin 5.5 Flash Preview: only 14 billion parameters are activated per token, and the weights are compressed using FP4 technology, which reduces the inference cost.
Regarding open-source: the model weights will be released to the global open-source community under the Apache 2.0 license before November 25, supporting SGLang and vLLM deployments. For developers, this means they can use the model immediately without any contractual obligations. Additionally, the API for Wenxin 5.5 Flash Preview is now available on the Qianfan platform.
@@It’s worth mentioning that, besides the actual tests and discussions about pricing, there is another piece of information that is often overlooked:
Some of the inference optimizations for Wenxin 5.5 Flash Preview were completed by Baidu’s automatic research system, ERNIE Evolve, which was announced on November 9. In low-latency scenarios using the Kunlun Core P800, ERNIE Evolve explored for approximately 168 hours, increasing the generation speed to 623.6% of the baseline, the throughput of deterministic inference to 133.7% of the baseline, and reducing the generation time per token by 26% without any loss in accuracy.
AI participating in its own optimization is still in its early stages, and a single optimization does not provide a complete picture. However, it indicates that the competition between models is shifting from a focus on parameter size to a focus on the efficiency of research and development iterations.
@@04
In conclusion, with three major actions within a week and a very competitive pricing strategy, Baidu’s approach with Wenxin is becoming clear: they are avoiding direct competition and instead focusing on efficiency, pricing, and open-source development simultaneously.It is obvious that this year’s model competition is becoming more segmented: the flagship models are competing for the top positions on the rankings, while the models in the efficiency category are competing on cost. Wenxin has chosen the latter approach. This segment of the competition is particularly crowded this year; pricing, demonstrations, and single tests merely serve as entry tickets. Whether these can be converted into actual usage by developers will determine the next stage of the competition. Another noteworthy industry trend is that video editing and game development are becoming new platforms for model deployment. These tasks do not rely on benchmarking results, and their outputs can be integrated into work processes, thus forming a second front outside of the traditional rankings. By focusing on these two areas, Wenxin is aligning itself with the industry’s development pace and making itself more visible to the market. Looking at the longer-term timeline, the period of pressure on Wenxin’s reputation coincides with a time when the company has been making silent progress and working hard to catch up. The industry still has high expectations for Baidu Wenxin, hoping that it will catch up with its own high-level technical capabilities. What kind of results will it deliver after this long period of development? AI Technology Reviews will continue to follow this closely. This article is originally from Leifeng Network and may not be reproduced without authorization. For more information, please refer to the reprint guidelines at and .
