TMTPost Hands-On Test of Baidu ERNIE 5.5 Flash Preview: Coding Capability Greatly Improved, Open-Sourced on November 25
Baidu released its foundation model ERNIE 5.5 Flash Preview at the World Conference 2026. It adopts a sparse MoE architecture, with 280 billion total parameters and 14 billion activated, and supports a one-million-token long context. TMTPost conducted hands-on tests with four types of tasks: interactive web pages, game development, video production, and research report writing. Among them, generating an interactive page from a single sentence took 136.88 seconds and about 46,000 tokens, and a 32-page storage industry research report was completed in under 5 minutes. Official API pricing is 2 yuan per million input tokens during peak hours and 1 yuan during off-peak hours, and 8 yuan per million output tokens during peak hours and 4 yuan during off-peak hours. The model weights will be open-sourced under the Apache License 2.0 on November 25, and support deployment with SGLang and vLLM.
Why read this
SelectedAI score 77
The author uses four types of hands-on tasks—web pages, games, videos, and research reports—to demonstrate this model's actual delivery level in coding and long context.
Topics
Body · AI translation
At the Baidu World Conference 2026, which opened on November 12, this giant, which once pioneered domestic AI but has also seen some ups and downs, released a series of cutting-edge new products. Among them, the most eye-catching is the latest foundation model ERNIE 5.5 Flash Preview. Sun Tianxiang, head of Baidu's Foundation Model R&D Department, also made his first public appearance to introduce the parameter scale and core optimization directions of this new model.
According to officially announced data, ERNIE 5.5 Flash Preview adopts a sparse MoE architecture, with a total parameter count of 280 billion and 14 billion activated, and supports a long context of one million tokens. For reference, these core parameters are roughly comparable to DeepSeek-V4-Flash.
In terms of capability focus, Baidu has also returned to the main battlefield. This new model performs prominently in programming, 3D modeling, knowledge work, content creation, and other areas—exactly the directions in which top models at home and abroad are placing their bets. Moreover, ERNIE 5.5 Flash Preview is also highly competitive in performance this time, achieving competitive scores in multiple benchmarks such as Terminal-Bench and AutomationBench, while maintaining extremely fast decoding speed.
Especially in programming capability—the "must-win battleground for models"—ERNIE 5.5 Flash Preview has improved substantially, scoring 1668 on the latest CodeArena: WebDev evaluation leaderboard, ranking second among open-source models and fourth among domestic models, closely following large-parameter domestic models such as Kimi K3.1 Pro and Qwen3.8-Max.
However, for Flash models, high cost-effectiveness and high execution efficiency are more important than paper parameters. In other words, the model must "do the work fast, well, and economically." In this regard, ERNIE 5.5 Flash Preview adopts a brand-new architecture design. Through multiple technological innovations such as hybrid sparse attention, multi-router MoE, and ECHO multi-token prediction, it improves model capability while reducing computation and memory overhead, making the processing of million-token long contexts and multi-step Agent tasks more efficient, thereby balancing intelligence level and inference efficiency. Data shows that its inference speed exceeds 200 tokens/s, and among open-source models of the same parameter scale, not many can stably achieve three-digit generation speeds.
In terms of pricing, ERNIE 5.5 Flash Preview is also very competitive. According to the officially announced API pricing, input is 2 yuan per million tokens during peak hours and 1 yuan during off-peak hours, while output is 8 yuan during peak hours and 4 yuan during off-peak hours. There is currently a limited-time 50% discount promotion. In addition, Baidu also announced that the model weights will be officially released to the global open-source community on November 25, and it has chosen the most commercial-friendly open-source approach—opening under the Apache License 2.0, under which developers can freely use, modify, and distribute it, and enterprises face almost no licensing concerns when integrating it. At the same time, it supports deployment through the two mainstream inference frameworks SGLang and vLLM, lowering both the barrier to access and migration costs. In addition, Sun Tianxiang said on stage at the conference, "This model name is still a bit long. We are working hard to remove the Preview suffix as soon as possible." It seems that the official version of the model is not far away.
Of course, the ultimate test for a model lies in real tasks and user experiences. Under the major trend of AI landing and forming productivity, only models that can deliver results and gain recognition have a future. We also selected high-frequency application scenarios that can demonstrate capabilities such as programming and knowledge work—including web interaction, game development, video production, and research report writing—and used real tasks to test the actual performance of ERNIE 5.5 Flash Preview.
Test 1 | Interactive page: ultra-fast delivery, with aesthetics under control
Our hands-on test started with a relatively simple requirement: generating an interactive webpage from a single sentence.
Recently, "After Meeting Friends" has become a hit and gone viral, so we planned to have ERNIE 5.5 Flash Preview ride the trend and create an interactive page for a "friendship personality test." The prompt was less than 200 words: create a "Friendship Personality Test" page where users answer 6 relationship scenario multiple-choice questions (such as "Would you take the initiative to speak after losing contact with an old friend?" and "Would you mind if the other person has a better new friend?"), with each question's options mapped to three dimensions: "taking the initiative to speak / clarifying boundaries / accepting fluidity"; after submission, calculate the score and generate a result character name and an interpretation; the result card includes the test name, score bars for each dimension, and a "Li Na style" quote, and supports one-click copying of the result text, retesting, an empty state, and responsive layout.
According to Baidu's official statement, the new model's programming capability has been greatly optimized, so this type of task is the most intuitive way to see the results. The output delivered by ERNIE 5.5 Flash Preview indeed has many impressive aspects.
First is the quality of the UI, which does not have the "plastic feel" common in AI-generated products on the market. Considering that it was generated from a single sentence, the model was able to choose this grid layout, minimalist Morandi color palette, and retro archive-folder style on its own—its aesthetic sense is indeed excellent.
Second, the interactive experience is also very smooth, and the functional completeness is high. From the first-screen onboarding to the quiz section and the results-sharing page, ERNIE 5.5 Flash Preview not only implemented the literal summarized requirements seamlessly, but also filled in on its own the progression logic and implementation details of an interactive quiz-style page, and the output results are also quite distinctive. After all, the ERNIE model once held the title of "strongest liberal arts student" and has always had its own unique strengths in "speaking like a human" and "providing emotional value." Now that its programming ability has improved, this "all-rounder" model has indeed delivered better results and also showcased its previous advantages and differentiating features in a more appropriate way.
Moreover, ERNIE 5.5 Flash Preview took only 136.88 seconds to complete the delivery, used about 46,000 tokens, and output nearly 350 tokens per second on average. In the comparative review, its output efficiency doubled compared with the previous generation, and its speed also has an advantage compared with mainstream models.
According to official disclosures, ERNIE 5.5 Flash Preview has made systematic front-end engineering optimizations around complex UI interaction, business logic implementation, state management, and full-stack integration, with the goal of making the generated interface both visually refined and functionally complete, and usable in production environments.
Judging from the delivery results of the interactive page, these efforts have already shown results.
Hands-on Test 2 | Game Development: It Can Write Mini-Games and Also Build Large Projects
The second type of task we gave was game development. According to officially announced information, ERNIE 5.5 Flash Preview has surpassed the initial stage of code completion and focuses more on real software engineering. It can fully build a wide range of coding scenarios for complex applications, and especially in long-cycle development tasks, it demonstrates stronger planning, execution, and continuous iteration capabilities. All of these have greatly improved the new model's game development results.
It seems that the official side also tends to believe that game development will be one of this model's important "stages for displaying its talents."
We also set up several tasks of different difficulty levels and types in this field.
First, we tried generating a 3D Gomoku game from a single sentence. ERNIE 5.5 Flash Preview delivered the result in less than 2 minutes. The aesthetics were still on point, the gameplay was also good, and the generation speed had an advantage compared with other mainstream domestic models. However, the ceiling for this type of "one-sentence generation" mini-game is indeed limited, and the effects and gameplay experience provided by different models are actually not much different.
Next, the task escalated: we asked ERNIE 5.5 Flash Preview to make a Monopoly-style simulation management game called "Fortune from All Directions." The prompt was still relatively short, and most elements were left for the AI to handle on its own. This time it took nearly 20 minutes. The game's complexity increased significantly, but it was not difficult to get started with, and the overall completeness was still very high. The settings for characters, industries, and shops were all very distinctive. In particular, it combined the randomness and contingency of this type of game with the strategy of simulation management well, enhancing the gaming experience.
Of course, with the rapid evolution of frontier models' programming capabilities, AI has long set its sights on more complex and larger games. ERNIE 5.5 Flash Preview is also said to have been deeply adapted for mainstream game engines such as Unreal Engine, Unity, and Godot, enabling the model to understand the development paradigms of different engines, skillfully use the corresponding tools and workflows, and complete complex development tasks from game creative design and logic implementation to scene construction. In an official demonstration, ERNIE 5.5 Flash Preview presented a very stunning case: the model spent 11 hours writing a third-person Eastern wuxia game called "Journey of Cloud Peaks" in C++ using Unreal Engine 5.8. The demo was very impressive. In particular, it is worth noting that, in a context where everything was generated solely through code, the game's camera logic was very reliable, and the character perspective, running direction, and smoothness of movement were also on point, with no obvious issues such as clipping, camera misalignment, or abnormal composition. Official information shows that although ERNIE 5.5 Flash Preview is a pure text model, it converts common visual understanding problems in game production into measurable code testing problems and, through continuously improved game testing and feedback, drives continuous iterative upgrades of the game until the goal is ultimately achieved.
Including ERNIE 5.5 Flash Preview, current frontier AI models, relying solely on their own programming capabilities, are already able to develop levels similar to those in AAA games from more than a decade ago. For the gaming industry, the development of AI is very likely to bring epoch-making changes.
Hands-on Test 3 | Video Production: Using Code to Complete the Closed Loop "From One Sentence to One Film"
We first gave ERNIE 5.5 Flash Preview a "touchstone task"—the currently very popular "15-second (Western) art history" video on social media. The prompt used was also a general English-language prompt, less than 100 English words. This time, ERNIE 5.5 Flash Preview took 40 minutes to "hand in the paper." The finished film linked key nodes such as prehistoric cave paintings, classical sculpture, Renaissance perspective, Impressionist light and color, and modern art within 15 seconds, achieving an effect comparable to the presentations by leading international models previously seen on social media.
Subsequently, we used a Chinese prompt of similar length to have ERNIE 5.5 Flash Preview make a version of a "15-second Chinese art history" video. The time spent generating the video was about the same, but the effect was even more stunning. From bronze patterns to ink-wash fine brushwork and then to modern style, the overall temperament was unified and the progression logic was reasonable.
After that, we also tried using ERNIE 5.5 Flash Preview to generate "a short video of dinosaurs doing the jump-flash dance," "an animated Odyssey trailer," and some video clips imitating the style of popular "AI short dramas."
Overall, the video quality generated by ERNIE 5.5 Flash Preview was consistently solid: the overall style matched the intent of the prompts, the script quality was especially impressive, and visually there were no cases of the image breaking down. Character and scene consistency was good, and camera cuts and action transitions were basically coherent. Even when character dialogue and actions were involved, there were none of the problems common to AI video, such as crossing the axis or continuity errors. Of course, in details such as the fineness of complex actions and the stability of long takes, this model still has room for further improvement.
In fact, what is most surprising is that ERNIE 5.5 Flash Preview completed the above tasks using only pure code, without calling external image libraries. We also read the relevant official release information in detail for this reason, and found that a notable feature of this model is that it has found a path to handle scriptwriting, storyboarding, resource lookup, and image generation and compositing through code and tool calls. What makes this possible mainly relies on two foundations: first, a full-process training environment covering creativity, scripts, storyboarding, materials, and voiceover, as well as more than 1,000 creation Skills; second, a million-token long context that can keep the file contents, tool call records, and returned results involved in the task within the window, allowing the model to call tools without "losing the thread."
It is precisely with these that a pure text model has run through the closed loop "from one sentence to one film."
Hands-on Test 4 | Knowledge Work: Completing an In-Depth Research Report in 5 Minutes, Scoring 93.8
We also tested office tasks such as PPT creation, spreadsheet organization, and meeting minutes, as well as some creative writing tasks. The results delivered by ERNIE 5.5 Flash Preview basically all reached a practically usable level.
In end-to-end knowledge work, the last topic we gave was writing an industry research report.
Recently, the storage sector has fluctuated sharply. We gave ERNIE 5.5 Flash Preview only the simplest requirement: given the assumed writer identity of "a semiconductor industry analyst with both industry analysis and equity research backgrounds," along with the research time point and core theme, we asked it to independently complete an in-depth research report on the storage industry.
In the end, ERNIE 5.5 Flash Preview completed a 32-page in-depth research report in less than 5 minutes, titled "Pricing Cracks in the Super Cycle."
We asked a third-party model to score this report and four other well-known storage industry research reports published by renowned institutions. ERNIE 5.5 Flash Preview's report received 93.8 points and, unexpectedly, became the only report to exceed 90 points, while the other reports scored between 81 and 88. The third-party model's explanation was: "Pricing Cracks in the Super Cycle" does not ask "is storage booming," but rather "why do record profits coexist with stock price pullbacks." It breaks the "super cycle" into three narratives: structural shortage, cyclical tightness, and transfer of pricing power, and quantifies price contribution, crowding-out coefficient, LTA volume locking without price locking, profit lock-in degree, depreciation wall, valuation cracks, and scenario observation indicators. Financial and risk penetration is the strongest, and the breakdown of definitions is especially professional. The deduction points lie in the fact that some calculations are subjective estimates, and there are places in the HBM definition reconciliation that cannot be fully verified.
Of course, AI evaluation dimensions also have their limitations, and newly generated reports often already "combine the strengths of existing reports," giving them a natural advantage in scoring. But either way, this result is enough to show that ERNIE 5.5 Flash Preview has a considerably high level of intelligence and delivery capability in the professional research field.
Betting on Programming and Self-Developed RSI: Baidu's Future Is Worth Anticipating
After hands-on tests across multiple tasks, ERNIE 5.5 Flash Preview gives a very distinct impression: whether in text, visuals, or user experience, its aesthetics and taste are consistently on point; the output content is concise and easy to use, approachable yet still refined.
In fact, when Baidu launched 5.0 Preview in November last year, its language capabilities, creative expression, and the cultural taste reflected in interactions attracted widespread attention. But at that time, these characteristics had not yet translated into advantages. One key reason was that there was a certain mismatch between the model's strengths and the user expectations and capability demonstration scenarios of that stage.
Late last year and early this year was precisely the moment when AI programming capabilities were moving toward large-scale adoption and Agent products were gathering momentum. Users were no longer satisfied with an AI that could only chat; they hoped it could truly integrate into their lives and work and become an intelligent assistant. So no matter how good a model was at chatting, it was hard for it to truly strike a chord with users.
Since then, Baidu's models seemed to go quiet for a while. But it was also during this period that the ERNIE model began to work hard to "catch up," with the focus on programming. By now, ERNIE 5.5 Flash Preview has delivered programming capabilities that put it among the leaders. Through its evolution in this field and levers such as Agent capabilities, it has driven scenarios where user demand is most concentrated, including web page creation, game development, 3D modeling, video generation, and knowledge work, allowing its former aesthetics and taste to bring out their differentiated advantages.
Breakthroughs on the model side are crucial for Baidu, which is doing its utmost to catch up and hopes to return to the center stage. At today's Baidu World 2026 conference, Baidu founder Robin Li also broke down the two development paths of the large model industry: one is to break through upward and push the upper limit of intelligence higher; the other is to take root in the real economy and unleash industrial value. He also stressed that Baidu is advancing on both fronts: on one hand, it will continue to invest in R&D of the next generation of more cutting-edge foundation models; on the other hand, it will adhere to the philosophy of "parallel models and agents, collaborative evolution," optimizing large models for applications.
In addition to the release of the ERNIE foundation model, Baidu also released two application models, the ERNIE Lightning deep-thinking model and the ERNIE Yijing micro-expression large model, targeting high-frequency complex tasks and high-emotion content production, respectively. According to the introduction, the ERNIE Lightning deep-thinking model is aimed at massive-request scenarios such as search, and through architectural innovation and a Looping scheme that coordinates fast and slow thinking, it has reached leading levels in public tests covering programming, reasoning, agents, and text. The ERNIE Yijing micro-expression large model targets video creation such as podcasts and short comic dramas, adopting unified modeling of text, vision, and speech, supporting 4,000-character text-driven generation and chapter-level dubbing, with video generation costs 50% lower than the industry, and its ultra-fast version can achieve 1:1 real-time generation. The fact that Baidu's foundation model and application models can "hand in their papers" simultaneously and achieve impressive results also shows that after anchoring its development direction, its internal synergies have begun to emerge.
In fact, Baidu has not only made up for lost ground and closed the gap, but has also charged into more cutting-edge fields and achieved something there. On November 9, Baidu ERNIE disclosed progress in the RSI (recursive self-improvement) field for the first time, unveiling ERNIE Evolve, an automated research system that participated in the R&D of ERNIE 5.5 Flash. According to the official introduction, it can, around goals and constraints, autonomously propose research directions, conduct experiments, verify results, and accumulate experience. It has achieved currently publicly verified SOTA performance on three evaluation metrics across the two tasks NanoChat AutoResearch and NanoGPT Speedrun, and has actually participated in the inference optimization of this new model. In other words, part of the capabilities of ERNIE 5.5 Flash Preview were researched by AI itself. RSI is one of the hottest topics and R&D directions in the model field today, and many believe it will completely transform the future of AI and even humanity.
English text is being prepared.