Baidu Releases ERNIE 5.5 Flash Preview, Focusing on Speed, Cost, and Open Source
Baidu has released ERNIE 5.5 Flash Preview, now available on the Qianfan platform, with both the web interface and API opened simultaneously; the model weights will be open-sourced under the Apache 2.0 license by November 25. The model has 280 billion total parameters and activates only 14 billion per token, and the Qianfan API delivers single-user inference speeds exceeding 200 tokens/s. 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; a limited-time 50% discount is currently available, and deployment via SGLang and vLLM is supported.
Why read this
AI score 72
The original text provides details on sparse activation, hybrid sparse attention, and pricing, which can be used to assess the practical trade-offs of an efficiency-tier model.
Topics
Body · AI translation
Today, Baidu's ERNIE 5.5 Flash Preview was officially released, marking ERNIE's third move within a single week.
On November 5, the base model ERNIE 5.5 Flash Base was open-sourced;
On November 9, RSI progress was disclosed for the first time, with the release of ERNIE Evolve;
On November 11, ERNIE 5.5 Flash Preview was released.
Three moves in a row within one week is rare in ERNIE's history. The three actions are closely linked, from the base model and RSI exploration to the Preview model, gradually revealing the new direction of Baidu's foundation model business.
In the long-standing external impression, ERNIE's position has been somewhat delicate: it started early enough and has accumulated a lot, yet its presence is not high enough. On the map of large-model competition, ERNIE has never quite met expectations.
In July this year, Sun Tianxiang, a 1997-born PhD from Fudan University, joined Baidu as head of the Foundation Model R&D Department (BMU) and also joined the Baidu Model Committee (BMC). He was once the leader of MOSS, one of China's first generation of ChatGPT-like open-source large models, and has also founded a startup. The outside world is awaiting his first report card after taking over the foundation model business. This week's triple strike may count as part of that report card.
English text is being prepared.
At tomorrow's Baidu World 2026, more information about the model will also be disclosed.
Let's start with the leaderboards. ERNIE 5.5 Flash Preview took first place in two of eight benchmark tests—AutomationBench and SpreadsheetBench V2—across Terminal-Bench 4.0, DeepSWE v1.1, AutomationBench, APEX-Agents 1.1, DRACO, SpreadsheetBench V2, OfficeQA Pro V2, and Agents' Last Exam; among the models included in the comparison, it ranked second in four tests—DeepSWE, APEX-Agents, DRACO, and OfficeQA Pro V2—and also placed near the top in the remaining items.
Baidu's official description of this result is that it "achieved competitive scores," a rather restrained choice of words.
It should be noted that scores marked with an asterisk in the comparison chart are based on vendors' self-reported metrics, and horizontal comparisons are for reference only.
Beyond leaderboard positions, the planning of the capability map and the engineering details in actual testing better demonstrate a model's true quality.
According to official statements, ERNIE 5.5 Flash Preview is upgraded mainly in four directions: coding, 3D modeling, knowledge work, and content creation. For coding, it was trained on more than 14 million executable code tasks and execution trajectories from multiple Agent frameworks; for 3D modeling, it received specialized training for tools such as Blender and CAD; post-training for knowledge work covered 300,000 workspaces and 10 million documents in fields such as finance and law; for content creation, it adapted to over a thousand Skills and built more than 150,000 creation environments, with official figures citing a cumulative total of about 5,300 minutes of finished footage.
Leiphone AI Tech Review tested ERNIE 5.5 Flash Preview's game development capabilities right away.
English text is being prepared.
The goal of the game is not to implement a program demo that can punch on the beat, but to create a complete, interactive, vibrant cartoon musical comedy.
The entire scene must feel like a genuinely existing absurd cartoon world: not only the player and the shipping robots, but also small characters constantly appearing around them, each with independent actions, reactions to the player's success or failure, a constantly changing environment, sudden intrusion events, and finally a climax performance in which all characters participate together.
The following is ERNIE 5.5 Flash Preview's performance in this test. The game successfully completed core segments such as characters gradually entering the scene, beat judgment, and settlement performances. After multiple rounds of generation, the total time was about 7 minutes and 32 seconds, ultimately producing a runnable version.
This test reflects to some extent ERNIE 5.5 Flash Preview's ability to handle highly composite requirements, while also examining the model's aesthetics, sense of musical rhythm, scene composition, and animation production capabilities; on the code side, it also required the correct triggering of a large number of events. Take the character state system as an example: if the state machine is poorly written, a character may be unable to recover from a tense state after consecutive failed hits, which can then affect chain reactions across the entire scene.
A single case cannot fully demonstrate the model's stable output level, but this type of long requirement task—"one document, multiple systems"—is itself one of the most demanding tests of coding ability.
The demo cases in the official blog cover areas such as video orchestration, game development, and 3D modeling, among which game development is the most substantial part.
Game development: self-correction
The representative case in game development is "Yunxiu Xing":
A third-person Eastern wuxia game set among the mist-shrouded stone peaks and cliff-top ancient temples of Zhangjiajie, built on Unreal Engine 5.8. The entire task took 11 hours; the model merged assets spanning multiple generations of the UE engine into a single UE5.8 project and fully implemented the game in C++.
Visual issues found during playtesting—such as characters floating in midair, inconsistent running directions, and misaligned camera positions—were each converted by the model into measurable code test problems, and the project was then iterated through increasingly complete test-driven development. "Translating visual problems into tests" means the model established a self-verification loop, which is closer to real development than one-shot generation of runnable code, and is precisely the hardest part to automate in long-cycle development tasks.
The official announcement also disclosed a larger-scale case: based on a complete game design plan, the model built nearly a hundred mobile mini-games, covering eight categories—bullet-hell shooting, roguelike, tower defense, racing, card RPG, city-building management, physics puzzle, and interactive narrative. Among them, 24 key works span multiple technology stacks including Cocos, Unity, LayaAir, PixiJS, and native Canvas 2D. Most use deterministic simulation with separation between the rules layer and the presentation layer, and automatically verify solvability and numerical balance through bot simulation, state-space search, and level replay—an engineering verification approach that is in the same vein as "Yunxiu Xing."
Video orchestration: rhythm, timing, and consistency
"Pelican Test" is a 117-second 2D animated short film:
ERNIE 5.5 Flash Preview wrote a 116-second 2D animated short film from scratch in Python. A white pelican rides a bicycle from the Sahara rock paintings of 4000 BC all the way into the code editor of the large-model era, traversing 20 art styles and finally arriving at the "Pelican Launch Event No. 21." The entire film is driven by tens of thousands of lines of code, with every frame vector-drawn using Pycairo and then post-processed with NumPy and Pillow. The same parameterized pelican transforms across 20 scenes into Han dynasty pictorial stone rubbings, Javanese shadow puppets, oscilloscope light traces, and 8-bit pixels; during scene changes, the image slides over as a whole frame like film, with the pelican always riding continuously across the seams of the frame.
English text is being prepared.
The thinking behind this task is worth pondering. In video generation competitions, single-frame image quality has already been pushed very high; the real difficulty lies in cross-scene consistency and rhythm control. The same pelican traversing 20 styles without losing its form demonstrates, to a certain extent, long-range state management capability.
In the 237-second official showcase film, the model, centered on the proposition "what if Earth also had a ring," continuously completed four processes: resource lookup, scriptwriting, storyboard design, and image generation. Official figures show that to complete one film, the model works continuously for one to two hours, with a median of about 170 tool calls, during which it checks results, finds problems, and corrects them on its own.
3D modeling: pipeline
Both 3D cases were completed based on Blender:
First, two agents based on ERNIE 5.5 Flash Preview used the Blender Python API in a division of labor: one was responsible for the airport, autonomously retrieving layout vector data from a map platform, downloading licensed assets such as plants, vehicles, and ground service equipment, and constructing the terminal and airfield through parametric modeling;
The other was responsible for the aircraft, constructing the fuselage and wings through cross-section lofting and modular assembly. The two projects were then unified in scale, position, and camera, animated, rendered frame-by-frame with Cycles/OptiX GPU, and composited into video with FFmpeg.
Second, the model autonomously retrieved references on the layout and architectural structure of Horyu-ji and gradually constructed the building complex, down to the details of tier-by-tier tapering, beam-and-column support, and layered roof tiles. Publicly licensed Buddha statue scans and plant assets were integrated into the scene, and surface details were completed jointly with PBR textures and procedural node materials.
The common point of the two cases is that the model was placed into a complete pipeline of "retrieval—modeling—rendering—compositing," rather than executing a single modeling instruction at one point. The task division and collaboration between the two agents also push the case beyond the scope of a single model's capabilities.
02
Where does speed come from
The game and video cases above are all long-horizon tasks: continuous work for one to two hours, a median of about 170 tool calls, and 11 hours of complete engineering iteration. Such tasks consume reasoning speed far beyond everyday Q&A, and "fast" has practical significance in them.
For most users, the more intuitive figure is this one: Qianfan API single-user inference speed exceeds 200 tokens/s, roughly equivalent to more than 150 Chinese characters per second. Once generation speed reaches this level, output can basically keep up with reading, and the sense of waiting drops significantly.
The official explanation of "fast" is a systematic technical design that can be viewed in three layers.
The first layer is how the parameters are used. ERNIE 5.5 Flash Preview has 280 billion total parameters, with only 14 billion activated per token. In current sparse configurations, this counts as a relatively aggressive tier. Total parameters determine the model's knowledge capacity, while activated parameters determine the actual computational cost per token. The 280 billion parameters are more like standing inventory; for each token generated, only a small portion is called upon.
The second layer is memory. If full attention were applied at every layer for a 1M-token context, the memory overhead would be too large to deploy. Hybrid sparse attention is the answer to this problem: of the 43 layers, 37 compress every 128 historical tokens into a single cache entry, preserving a global view; the remaining 6 layers retain the ability to access history token by token, combined with sparse selection. The effect is a substantial drop in KV cache usage while detail retrieval in long texts is still preserved.
The third layer is acceleration during inference. FP4 compresses the storage of expert weights to 4 bits, directly saving memory bandwidth; ECHO lets the final layer of the backbone network, in addition to predicting the next token, also predict subsequent tokens, with this layer simultaneously serving as the draft model for speculative decoding—guess first, verify later, gaining speed with almost no added parameters.
The three-layer design points to the same goal: deliver sufficient intelligence with fewer activated parameters, raise speed, and lower cost further. This is also the industry convention behind the name "Flash"—compared with pushing a single point to its limit, the core logic of the efficiency-tier product line emphasizes the balance of capability, speed, and cost.
On pricing, the API costs 2 yuan per million tokens for input during peak hours and 1 yuan during off-peak hours, and 8 yuan for output during peak hours and 4 yuan during off-peak hours, currently at a limited-time 50% discount—after the discount, off-peak input is 0.5 yuan and output is 2 yuan. Among models of the same parameter scale, this is a competitive price.
A conversion makes it more intuitive: one million tokens is roughly 600,000 to 800,000 Chinese characters. Taking the longest context the model supports as an example, filling an entire 1-million-token window costs less than one yuan in input after the discount.
The efficiency design opens up room for the low price of ERNIE 5.5 Flash Preview: only 14 billion parameters are activated per token, FP4 compresses expert weights, and inference cost has already been pushed down by the architecture.
On open source, the model weights will be released to the global open-source community under the Apache 2.0 license before November 25, supporting deployment with SGLang and vLLM. For developers, this means pull and run, with no licensing burden for commercial use. In addition, the ERNIE 5.5 Flash Preview API went live on the Qianfan platform today.
It is worth noting that, beyond the hands-on tests and heated discussion of pricing, there is a piece of news that is easy to overlook:
Part of the inference optimization work for ERNIE 5.5 Flash Preview was completed by ERNIE Evolve, the automated research system Baidu disclosed on November 9: in the low-latency scenario on Kunlun chips P800, ERNIE Evolve explored continuously for about 168 hours, raising generation speed to 623.6% of the baseline; deterministic inference throughput increased to 133.7% of the baseline, average generation time per token was shortened by 26%, and accuracy did not drop.
04
Conclusion
Three moves within a week, plus very sincere pricing, and the outline of Baidu ERNIE's latest offensive is already clear: bypass the frontal battlefield and first place pieces simultaneously on the three dimensions of efficiency, price, and open source.
Obviously, this year's model competition is stratifying: flagship models compete over the upper limits of leaderboards, efficiency-tier models compete over the cost curve, and ERNIE has chosen the latter. This track is especially crowded this year; pricing, demos, and single hands-on tests are only entry tickets, and whether they can translate into developer call volume is the dividing line for the next stage.
Another industry change worth recording is that video orchestration and game development are becoming new showcases for model releases. Such tasks do not depend on benchmark scores, their outputs can enter workflows, and they are forming a second front beyond the leaderboards. ERNIE placing its strengths in these two scenarios is playing along with the industry rhythm and also making itself most easily seen.
Taking a longer view of the timeline, this period of reputational pressure was also precisely a time when ERNIE was quietly making progress and catching up with a low profile. The industry still has expectations for Baidu ERNIE, expecting it to catch up with the version of itself that has rich technical accumulation. After a long dormant period, what kind of answer will it deliver next? AI Technology Review will continue to follow.