Water from the Temple of Penang
QbitAI is public.
Just now, Baidu has sent out a new model, and the netizens on X can't sit down to watch the demo.
Baidu, is this going to be supported?
It's old-fashioned to shout "Ride a Bicycle", but have you ever seen a version like this?
Concept design comes from the creator of Bilibili @Step CloseA monkey riding a bicycle, from a Sahara rock painting of 4000 BC, goes all the way to a code editor from the era of the Big Model.
116 seconds 20 different art styles, all in Python code.
All I can say is that the monkey is the monkey, but the game is no longer one-dimensional.
It's just an appetizer.
I just finished the animation, the model behind it.Turning his head into the U.E. 5.8 fantasy engine, he worked for 11 consecutive hours, producing a game of oriental martial arts.。
It's all set up, with the rock tops, the cliff tops, the characters, the game logic.
It's not like, guys, the front foot is code animation, the back foot is code game creation, what's the big trick this time?
The answer is:
This is the latest version of ERNIE 5.5 Flash Preview.
Just the day before the Baidu World Congress, the new model was officially unveiled and announced to be open source.
A total of 280 billion parameters, with a single reasoning to activate only 14 billion parameters, support the context of millions of tokens and generate more than 200 tokens per second.
This time around, the skill tree highlights only two words:
It's programming.
Note that programming is more than just writing web pages, adding code, game development, 3D modeling, video editing, which requires continuous calling tools, repeatedly testing and modifying complex tasks.
Currently, in programming and agent evaluations such as Terminal-Bench, AutomationBench and others, Centaur 5.5 Flash Preview has demonstrated the strength of competing with flagship models such as Claude Opus 5.5 and GPT-6.1 Sol.
The atmosphere is so hot here, it's like, "The chicken is out of the frying pan".
3D, games, web pages... Awareness 5.5 Flash Preview, please follow me.
Flash Preview of the 5th edition
Currently, both the web end and the API can experience this model.
The first question is about the role of the monkey in the monkey's sea turtle.
The nine-tailed fox.
But this time, I'm going to have literature code for an animation called "Bronze Organ Beast Awakening".
Starting with an ancient manuscript, the turrets, skeletons, and skeletons flew by themselves, eventually forming a nine-tailed organ animal.
It is completely forbidden to use external materials, modeling, animation, or rendering entirely on Python.
We'll see what comes out.
Hey, dude, who's the nine-tailed fox, figuring out if it's a cyber-organic beast that ran out of a mountain gorge?
What about video games?
The second question, which I decided to emphasize directly:It's a great place to play in the back room.
And it can't just be a yellow corridor that's messing around, the culture has to do a complete collection and transit process.
The player must find three video tapes in the labyrinth, encounter physical chase along the way, physical strength, failure and export mechanisms, and finally deliver a web page that can be opened directly in the browser.
As a result, Literature gave me a complete single HTML file.
You can play without installing a game engine or setting up an additional environment.
Yellow wallpaper, wet carpets, thundering sunlight, and an old video with a filter, all came out of nowhere.
In the beginning, I was quite relaxed.
It's all about going around the corridor and looking for videos.
And then...
One of the players was playing and suddenly he was caught on the spot.
That's good, man. I was frightened by the cold.
It's a little too fast, it's a little too shallow, and it's still a beautiful human world.
The third question, I threw literature directly into Blender and let it create a Suzuki garden from scratch.。
All building, plant, water and lens animations have to be built using the Blender API in Python.
Then, literacy began to build a garden in Blender.
After about half an hour, the model returned the following results:
In fact, I was a little surprised by the effect.
Anyway, this time it's a preliminary touch.
Officially, these complex demos run for ten hours, with no more detailed task design, tool collaboration and repetitive debugging.
Where the limits of Flash Preview 5.5 are, we have to keep trying.
But at the end of this cycle, I was curious about something else:
It's the same thing with coding, and how does the literacy go from understanding the needs to the "end-to-end links" of project delivery?
The answer may lie in a capacity that even Jeff Dean has underestimated.
Jeff acknowledges that the "code" is in the eye of the beholder, adding to the literacy.
Recently, in a conversation with Lee Pfeiffer, Jeff Dean proactively admitted that he was overlooking one thing.
Programming。
Gemini also practiced coding in his early years, but did not put it at its core until he later discovered that the impact of programming on the model to solve various problems was much deeper than he imagined.
Let's put it this way, programming is like a pair of hands growing out of a chat box, with its model to get away from the category of "paper talkers".
But as time goes on, everyone's expectations for programming are getting higher and higher.
Sources: Oil pipeline account @stanfordhaiNow that AI has cracked the code, it's making a web page, fixing a bug, and giving results in a few minutes.
But to do a project independently, trouble comes.
The need for demolition, environmental configuration, testing and modeling, no one can stop staring.
Well, guys, let's just say that AI writes code, so humans have to go all the way.
So the ideal coding should be:
It's just a matter of getting it in and out.
You tell the AI what it wants to do, dismantle tasks, write code, call tools, test, fix bugs, and it does it by itself, step by step.
This is the so-called "end-to-end" and the ability of literature to emphasize this.
So what exactly is this? Let's see a number:
14 million.
This is the number of executable code tasks that Baidu is preparing for Flash Preview 5.5.
It's a big number, but it's more important."Executable"。
Each code is thrown into a real environment and the task of judging the results is completed.
The code is beautifully written, and it can't run and it's white.
Do you want to write it down?
Sorry, it's not that easy.
As for how to find the problem after the code runs, then change it, Baidu also uses multiple Agent execution tracks to teach the model when to use the tool, how to view feedback, what went wrong and how to fix it.
It's easier to understand if you put it in a grasshopper.
In the process of development, the model encountered problems such as suspended characters, incorrect driving directions, and misplaced camera positions.
But the Flash Preview 5.5 is a text-only model, and you can't look directly at the game screen for Bugs.
How about that?
Baidu converted these visual problems into code-checkable data and then continued to modify the model based on the test results.
It's like running again, checking the problem, changing the code, and then running again.
The modeling process, which originally required developers to watch the screen finish, was also embedded in the model's own work cycle.
Of course, learning the Loop doesn't mean you can do anything alive.
Web pages, games, 3D modeling, and video creation all look like they're writing code, and the gateways are completely different.
So Baidu is back around the specific scene, giving the model a lesson by the lesson.。
To make a game, you have to understand the game's mechanics and how it works, and you have to adapt to Unreal Engine, Unity, Godot, etc.
When it comes to blenders and CAD, you have to learn how to use professional tools to run a whole set of modeling processes continuously.
It's more difficult to create videos.
From scripts, mirrors, audio, and clips, all tools have to be rotated in between, and Baidu has prepared more than 1,000 Skills for this purpose.
It's not easy to teach tools, and a new problem arises:
The memory.
It's easy for an agent to work for hours at a time and forget about "before school".
So the Climax 5.5 Flash Preview directly pulls the context up.One million tokensIt can accommodate more project documents, code and tool call records.
There is not enough light to remember, but critical moments are still to be found.
Literature also uses mixed scarcity of attention to help models quickly extract useful information from a wealth of historical records.
It combines technologies such as multi-router MoE, ECHO Multi-Token prediction and FP4 expert weights, ultimately pushing the generation speed to over 200 tokens per second.
It should be noted that AI is already involved in this optimization work.
For the first time, Flash Preview introduces self-developed RSI technology, which is being developed by ERNIE Evolve to engage in model reasoning optimization and explore new pathways to AI involvement in model research and optimization.。
At the end of the day, the hardest thing to do when you want AI to finish a project is two things:
First, try not to get caught in the middle.
Secondly, there must be sufficient durability. The longer the child lives, the faster the memory, context and production speed of the model is tested.
Flash Preview 5.5 is an enhanced version of Coding, and it's aimed at both of them.
From a piece of code that can be seen, to a project that can really be used, the middle path that can't be completed on its own is the real content of "end-to-end".
One More Thing
It's been a while since I've been able to talk about this, and it's always felt like:
This team seems a little different.
It's not about a conference or a benchmark list.
It was a coincidence.
A few days ago, after reading a book, Junyuan Shang, a member of the ERNIE Pre-Training Team, posted a video of his handshake.
The cover is only six syllables.
I'm still a teenager.
It's a little bit of a middle ground.
But with the phrase "a new beginning for a new generation, let's stay fresh, always with love, always moving forward", recalling the experience of these years in Baidu literature is suddenly a bit unpleasant.
Baidu is one of the first companies in the country to come out to make big models.
But the model generation has been updated and the outside world has not stopped questioning.
Especially the phrase, "Get up early, catch up late", which the team is expected to hear very quickly.
The team's atmosphere began to change.
I noticed that literature center researchers started coming to the front of the stage, actively interacting with netizens on a little book, X.
Even in their past disputes and complaints, they began to face each other.
The wave was a "collective business" that even caught the eye of netizens:
Why did the ERNIE team come to X?
I'm not sure if there are any indicators.
But from one person to a group of people, some of the researchers who were hiding behind the model's name, are actually starting to get in front of the stage.
It's not easy to get caught up in a spit, and to get caught up in a problem is to be willing to respond positively.
It is clear from these changes that there is a catch:
This guy, it seems, is also holding his breath, trying to make things right.
However, with the attitude, "Boooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooooo
Don't tell me, Baidu is real.
Let's do two things:
One is the announcement of an open-source plan, and the other is a limited time discount.
The Flash Preview 5.5 is expected to be released in the near future under the Apache 2.0 license and support SGLang and vLLM deployments.
There's nothing to hide here.
Developers want to deploy, re-develop, or directly pull down to find problems.
In terms of price, the model API is also synchronized to the Baidu Intelligent Cloud Sail platform, which currently has a five-fold limit (input of 1 yuan/million tokens, output of 4 yuan/million tokens, twice as much when busy).
It's really important.
In the past, games, 3D and video tasks were often more expensive than tokens, and even more powerful models were unusable.
"Push one", the road is paved and the rest is left to the developers.
If the model is not able to play, then everyone will answer for themselves.
In the video above, the phrase "returning is still a teenager" comes to mind.
After leaving so early that year, he was spoken of as "haunting a night party".
Today, even this group of models, with a few points of dissatisfaction, are looking forward to pursuing it further.
As for what we can't do this time, we can't do something different.
Then... wait and see.
It's the door.
