Baidu Releases ERNIE 5.5 Flash Preview, Focusing on End-to-End Programming and Announcing Imminent Open-Sourcing

Just now, Baidu released a new model, and netizens on X couldn't sit still after seeing the demo.

Is Baidu about to rise up??

Everyone's saying "a pelican riding a bicycle" is already old and outdated, but have you ever seen a version like this?

English text is being prepared.

A pelican riding a bicycle, from 4000 BC Saharan rock art, all the way into the code editor of the large model era.

116 seconds, 20 artistic styles, with the entire animation orchestrated through Python code.

All I can say is, the pelican is still that same pelican, but the way it's played is no longer in the same dimension...

And this is just an appetizer.

Right after finishing the animation, the model behind it turned around and dove into Unreal Engine UE5.8, working continuously for 11 hours to crank out an Eastern wuxia game, "Cloud Peak Journey."

Zhangjiajie stone peaks, ancient temples atop cliffs, character movements, game logic... it actually got all of it done.

No way, bro—first using code to shoot an animation, then using code to build a game right after. What big move has Baidu been holding back this time??

The answer is revealed:

ERNIE 5.5 Flash Preview.

Just one day before the Baidu World Conference, this new model officially debuted, with an announcement that it will soon be open-sourced.

280 billion total parameters, only 14 billion activated per inference, support for a million-token context, and a generation speed exceeding 200 tokens per second.

And this time, the skill tree ERNIE is highlighting comes down to just two words:

Programming.

Note that programming here is no longer just about writing web pages or filling in code. Complex tasks that require continuously calling tools and repeated testing and modification—game development, 3D modeling, video orchestration—the model can also handle directly.

Currently, in programming and Agent benchmarks such as Terminal-Bench and AutomationBench, ERNIE 5.5 Flash Preview has already shown the strength to compete on the same stage as flagship models like Claude Opus 5.5 and GPT-6.1 Sol.

The atmosphere has been built up this far, so let's "take the mule or the horse out for a walk."

3D, games, web pages... ERNIE 5.5 Flash Preview, bring it on. (doge)

But this time, I want ERNIE to use code to orchestrate an animation of the "Awakening of the Bronze Mechanical Beast."

Starting from an ancient manuscript, gears, armor plates, and skeletons must fly up on their own, and finally assemble into a nine-tailed mechanical beast.

No external assets allowed throughout; modeling, animation, and rendering all rely on Python.

The result came out, and looking at it—

Good heavens, this is no nine-tailed fox; it's clearly a cyber mechanical beast that ran out of the Classic of Mountains and Seas?!

If it can shoot videos, then what about games?

For the second task, I decided to directly raise the intensity: crank out a playable "Backrooms."

And it can't just build one yellow corridor to fool people—ERNIE must create a complete collection and clear-the-game flow—

Players need to find three videotapes in the maze, encounter entity chases along the way, and there are stamina, failure, and exit mechanics. The final deliverable must be a web page that can be opened directly in a browser.

As a result, ERNIE served me a complete single HTML file.

No need to install a game engine, no need for extra environment configuration—just click and play.

Yellow wallpaper, damp carpet, buzzing fluorescent lights, plus an old videotape filter, and that vibe comes out immediately.

My heart was beating too fast; it felt a bit too underworld, so let's go for something from the beautiful human world instead.

For the third task, I threw ERNIE directly into Blender and had it build a Suzhou garden from scratch.

All buildings, plants, water surfaces, and camera animations had to be built bit by bit by calling the Blender API through Python.

Then ERNIE started huffing and puffing away in Blender, building the garden.

About half an hour later, the model returned the following "one-shot" result to me.

To be honest, this effect really did surprise me a bit...

Anyway, this hands-on session is at most just a preliminary probe.

Those official complex demos that run for a dozen-plus hours inevitably involve more detailed task design, tool coordination, and repeated debugging.

Where exactly the upper limit of ERNIE 5.5 Flash Preview's capabilities lies still needs further testing.

But after running through this round, I'm very curious about another thing:

Both are Coding—how did ERNIE this time open up the "end-to-end link" from requirement understanding to project delivery?

The answer may be hidden in a capability that even Jeff Dean once underestimated.

Jeff admits he misjudged "code," and ERNIE has focused on filling this gap this time

Not long ago, in a conversation with Fei-Fei Li, Jeff Dean proactively admitted that he had misjudged one thing—

Programming.

Gemini was also practicing Coding early on, but did not put it at the core position. Later he discovered that programming ability's impact on a model's ability to solve various problems is far deeper than imagined.

Put it this way: programming is like a pair of hands growing out of the chat box. With it, the model can escape the realm of "armchair strategizing."

But the further we go, the higher everyone's expectations for programming become.

△Image source: YouTube account @stanfordhai

Now AI writing code is already quite smooth—making a web page or fixing a bug can give results in a few minutes.

But when it comes to actually completing a project independently, trouble arises.

Requirement breakdown, environment setup, testing and debugging—every step needs someone watching over it.

So the ideal Coding should be:

Requirements go in, a finished product comes out.

You just tell the AI what you want to do, and it breaks down tasks, writes code, calls tools, tests, and fixes bugs, all on its own, step by step.

This is the so-called "end-to-end," and it's also the capability ERNIE focused on strengthening this time.

How exactly did it strengthen it? First, look at one number:

14 million.

This is the number of executable code tasks Baidu prepared for ERNIE 5.5 Flash Preview.

The number is huge, but what's more critical is "executable."

Every piece of code has to be thrown into a real environment to run, and then the results are used to judge whether the task was completed.

As for how to find problems after the code runs and then keep fixing them, Baidu also used various Agent execution trajectories to teach the model when to use tools, how to read feedback, and how to fix errors when they occur.

Put into the context of "Yunxiu Xing," it becomes easier to understand.

During development, the model encountered problems such as characters floating in the air, running in the wrong direction, and camera positions being messed up.

But ERNIE 5.5 Flash Preview is a pure text model and cannot directly stare at the game screen to find bugs.

What to do?

Baidu converted these visual problems into data that code can check, and then used the test results to have the model continue making modifications.

Just like that: run it once, check for problems, modify the code, then run it again.

The debugging process that originally required developers to stare at the screen was also stuffed into the model's own work loop.

Of course, learning this Loop doesn't mean it can do any kind of work.

Web pages, games, 3D modeling, and video creation all look like writing code, but the know-how inside is completely different.

So Baidu then gave the model remedial lessons one by one around specific scenarios.

For games, it needs to understand gameplay mechanics and runtime states, and also adapt to engines such as Unreal Engine, Unity, and Godot.

For Blender and CAD, it needs to learn to call professional tools and run an entire modeling workflow continuously from start to finish.

Video creation is even more troublesome.

From scripts and storyboards to voiceover and editing, various tools have to take turns passing the baton in between, and Baidu prepared more than 1,000 creation Skills for this.

After finally teaching it the tools, a new problem emerged:

Memory.

When an Agent works continuously for several hours, it easily "forgets what it learned earlier" as it goes.

So ERNIE 5.5 Flash Preview directly pushed the context to 1 million Tokens, able to hold more project files, code, and tool-call records.

Just remembering a lot is not enough; at critical moments it also has to be able to find things.

ERNIE also used hybrid sparse attention to help the model quickly dig out the information that is truly useful right now from a large pile of historical records.

Combined with technologies such as multi-router MoE, ECHO multi-token prediction, and FP4 expert weights, it ultimately pushed generation speed to over 200 Tokens per second.

ERNIE 5.5 Flash Preview introduces its self-developed RSI technology for the first time, with ERNIE Evolve participating in model inference optimization, exploring a new path for AI to participate in model research and optimization.

First, try not to have a human take over midway.

Second, it needs enough endurance. The longer the work goes on, the more it tests the model's memory, context, and generation speed.

ERNIE 5.5 Flash Preview's strengthening of Coding this time targets exactly these two hurdles.

From a piece of code that looks okay to a project that is truly usable, whether it can walk that entire path by itself is the real value of "end-to-end."

It's not because of some launch event, or some Benchmark leaderboard.

The reason was actually quite accidental.

A few days ago, while scrolling through Xiaomou Shu, I saw a video hand-made by Junyuan Shang from Baidu ERNIE's pretraining team.

But paired with that line "A new starting point for reshaping a generation—may we keep our original aspiration, always hold love, and keep moving forward"—and then thinking about what Baidu ERNIE has been through over the years, I suddenly felt a bit uneasy.

Baidu counts as one of the earliest companies in China to enter large models; it didn't start late, and it has quite a lot of technical accumulation.

But as the models updated generation after generation, outside doubts never stopped along the way.

Especially that line "rose early but arrived at the market late"—the team itself has probably heard it so much they've grown calluses.

As they kept hearing it, the team's style gradually began to change...

I noticed that ERNIE R&D staff began stepping into the foreground one after another, proactively communicating with netizens on Xiaomou Shu and X.

But from one person to a group of people, some R&D staff who were originally hidden behind model names have indeed begun to step into the foreground.

Still, attitude is attitude—"words of concern are not as good as a big payment"; you have to let people see concrete action.

The model weights of ERNIE 5.5 Flash Preview are expected to be open-sourced soon, under the Apache 2.0 license, and will support SGLang and vLLM deployment.

Now there's nothing left to hide.

Developers who want to deploy it themselves, do secondary development, or just pull it down to look for problems can all do so.

On pricing, the model API has also launched simultaneously on Baidu AI Cloud's Qianfan platform, currently at a limited-time 50% discount (1 yuan per million input Tokens, 4 yuan per million output Tokens; both double during busy hours).

This point is actually quite important.