This model is fast, cheap, open source.
It's fast, it's really fast. It spews more than 200 tokens, 280 billion parameters, and only moves 14 billion of them every second.
It's cheap, it's really cheap. API for every million tokens, 2 yuan per hour of input, 1 yuan per hour of output, 8 yuan per hour of output, 4 yuan per hour of output.
Baidu's Intelligent Cloud Seals, which can be directly tuned by the API, will go online today; the model rights will be released to the global open source community on November 25.
5.5 Flash Preview is fully upgraded for programming, 3D modeling, knowledge work, and content creation.
In several benchmark tests, such as Terminal-Bench, AutomationBench, it yielded competitive scores:
Terminal-Bench 4.0, with a real gun in the terminal, scored 46.5; AutomationBench 61.9; and SpreadsheetBench V2, with a real Excel, scored 36.4.
In the graph, a * is a graded self-determination score.
There is also a project that is still being pursued, but don't forget its location: a fast-paced and saving Flash model.
280 billion parameters, only 14 billion tokens activated at a time; 1 million tokens in context; more than 200 tokens can be thrown out per second.
In this position, the top two finishers, along with Claude Opus 5.5 and GPT-6.1 Sol, also took the podium.
When I run this thing, I only trust half of it. The other half, I have to see it work.
The first thing I'm going to do is take a look at this:
Unreal Engine 5.8 is a third-person shooter game called Eastern Man. The game's engine is Unreal Engine 5.8, and the underlying code is C++, which was developed by Visual Studio 5.5 Flash Preview in an 11-hour mission.
11 hours is not uncommon. What is rare is that the 5.5 Flash Preview is a pure-text model that cannot be seen in images.
One of the most common problems in 3D games is that the characters are suspended in the middle of nowhere, they run and stare in the direction of the camera.
It's a very sharp and clever solution:I can't see it, that's the size.
For example, a person hangs in the air, without looking at the picture, writing a paragraph, and you know how high the program is at the bottom of the foot from the ground. It takes these visual problems that arise in the game, a change in the code to run, test the number of code, then measure, change, measure again, a round of grinding down until the game is complete.
In the meantime, it did a dirty job: combining material from generations of fantasy engines into the same UE 5.8 engine. Different generations of engines, formats and interfaces are often incompatible.
To tell you the truth, the case that struck me was not the picture, but the way it worked.
Anyone who has written the code knows that the beginner bug says "it's not right" and the boss says "the return value on the 37th line is 0.3".
From a mouse to a W16:The code is its omnipotent key.Where did the endurance come from, working for 11 hours straight without collapsing?
According to official disclosures, it used more than 14 million "true run-of-the-mill" code missions during its training, plus the execution trajectory of multiple Intelligent Agency frameworks.
It's a whole set of "write, run, fix, then change" moves that don't use back-end codes.
In addition to the over a million tokens in the background, the documents read in eleven hours, the tools transferred, the results obtained are all in your mind. Unreal, Unity, Godot, the mainstream game engines, are also specially adapted.
We also came up with a question: how to make a third-person 3D swordflight game, swordfighting through clouds.
The main character is dressed in a blue robe, with a light-colored ribbon on his back, with a haircut on his head, steadily stepping on a flying sword, without taking several square brackets.
Looking ahead, there is a blue-green stone peak in the warm golden cloud sea, a stone in the middle of the path, and ten golden rings that line the temple to the top of the cliff.
The key is that it really plays through. Flying through the 3rd and 7th circles, check points automatically note; the number of circles, timing, energy bars in the upper left corner continue to run, and there is a small map in the lower right corner. Finally, in front of the red column-heavy temple, the scoreboard pops out: ten circles complete.
All right, let's get it to make a Copper film. The title is just one sentence: What would the world look like if the moon was only halfway away from Earth?
The film takes place over a period of more than a minute, screenshots. At the beginning of a huge circle of moon pressed over the surface of the sea, the phrase "the first thing that gets out of control is not the night sky, it's the sea".
I was surprised to learn that it was partitioned. Speaking of tides, it specifically added a sentence: it is eight times the tidal force, not eight times the tidal height. No one asked, it itself calculated the Loshi limit, telling you that the moon will not be torn, and that there is more than ten times the distance from that line.
Some of the key numbers I've run through with my calculator, and they're all right. One Flash model is the life of an entire group of Cope videos: someone checks the data, someone calculates, someone draws a map, someone cuts a clip.
It's the same story, the same place, the same spirit.
In the developers' circle, there was a widely circulated soil taste test: the model had to use code to draw a bicycle saddle.
A picture was handed over by someone else. A picture was handed over by Visual 5.5 Flash Preview.
A white-collared bicycle, starting with a Sahara rock painting of 4000 B.C.E., riding a Han portrait, a Javanese sculpture, an 8-bit pixel, changing 20 artistic styles, constantly riding the code editor of the Big Model era, and finally climbing the "21st edition".
The soundtrack features 22 tracks, ranging from drumming, guitars to Mario themes. How to cut is not for the faint of heart: the model himself wrote the scripts for the clips and aligned each of the alternate spots.
The title of the episode is just one monkey, and it gives a brief history of monkey.
It's been practiced.
The official disclosure is that these scenes revolve around KOP, comics, creative clippings during training, with more than 150,000 creative environments, ranging from creative, script, lens to a soundtrack, with a cumulative film output of about 5,300 minutes.
It's about spatial perception: given only a word to describe, it has to infer its own shape, proportions, and position relationships.
In the 3D modeling software Blender, two artificial intelligence units are in operation, one to build an airport and one to build an aircraft.
The airport building itself goes to the map platform to find layout data, download vegetation, vehicles, ground handling equipment and other supporting materials; the aircraft builder assembles a portion of the body, wings, cockpit, double-decker cabin.
In engineering design software FreeCAD, it uses a W16 engine with a Python script, which rotates the axis, and a four-row piston engine follows.
It also built two-story, editable three-dimensional models similar to the Baidu International Dam in Shenzhen, localized to the interior layout and power lines; and drew inspiration from Yuuyang II, designing a lunar concept car with a suspended, rotating sun wing.
This is where I come back from.
Games, animations, airplanes, engines, watching chatter play no more, the bottom line is the same: writing code.
Modeling is where it writes scripts and commands Blender; cutting is where it writes programs to coordinate audio and video material.
So Baidu listed programming skills as the number one highlight of this upgrade, and I'm convinced.
This is the key to the opening of the office door. The opening of the first two, the examination is the kind of living children.
It independently wrote a 24-page self-driving research report, completed 4169 U.S. Federal Accident Reports, and unlocked the safety data on both Waymo and Tesla routes. It also created a personal wealth planning tool that simulates 1000 random market routes that you can't afford to pay for retirement.
 |  |
This is also practiced. For financial, legal and other knowledge-based work, it has been officially given 300,000 workspaces, 10 million references and covers more than 40 types of documents.
He even composed it himself, producing a 6-minute, 35-second symphony.
This "quick" one.Part of it is AI itself.To be able to do one thing, to do it quickly, to be able to do it is another thing.
The smart-body activities take hundreds of steps, one step at a time, one step at a time, and the bills at the end of the month are heartbreaking.
Flash Preview 5.5 The two words "efficiency" are deadly in architectural design:
Read the text and save itThe model consists of 43 layers, of which 37 layers only look at "abstracts" when reading historical content, each comprising 128 tokens; the remaining 6 layers are word-for-word translations, and only select the most critical locations. Like a project group, most people look at the minutes of the meeting and a few people are responsible for the original archives.
Choose experts to keep an eye out for.There are 512 "experts" at each level, with only 10 points each time, plus 1 resident sharing experts, 2800 billion parameters that are compressed into just 14 billion each time.
Writing a draft takes timeThere is also a design called ECHO: without extra modules, directly using the last layer of the model itself, when writing a word, write a draft of the last few words by hand to verify the correctness.
There are several conflicting components on the architecture: the mHC enlarges the information channel between the layers and the layers, the MultiMax pipeline picks up the last word that step, and the R2L denominator cuts the number from right to left in a set of three digits, making it convenient to calculate.
When combined, more than 200 tokens are generated per second.
It's more interesting in the back.
On 9 November, Baidu was the first to reveal the progress in the direction of RSI: an automated research system called ERNIE Evolve, which allows AI to be confused when researchers, who come up with their own ideas, change their own code, run their own experiments, and have independently evaluated the results.
It was given a real-world mission to speed up the logic of the 5.5 Flash.
In the low latency scenario of the domestic cube core P800, the generating speed refers to 623.6% of the baseline. In the "certainty reasoning" used in reinforcement learning training, the same request, each word generated, each probability scored no less, the cost was slowed down), the throughput increased by 33.7%.
This scenario has been specially optimized by human engineers, and easy-to-get gains have long been dried up, with the remaining single improvements often being only 0.3% to 0.5%, smaller than measuring noise.
It tried hard in seven directions and squeezed 14.9% of the speed.
AI is harnessing the next generation of AI. This phrase used to live in science fiction and is now recorded in a research and development record of a large domestic model.
Seven Days of Silence: Catching up with the lostIf you look back on the calendar, the pace of the seven days is a little intimidating.
On November 5th, Windows 5.5 Flash Base is open source, which is 16 months after Baidu, releasing another model weight of 100 billion parameters. On November 9th, RSI reveals the progress of automated research.
The next chapter is set for release on November 25th, with a 5.5 Flash Preview powered open source.
And tomorrow, Baidu World 2026 will open in Beijing, with the theme of "The next act, intelligent organization".
It's the truth.
For the past two years, there has been a lot of talk about literacy, expectations and questions.
But what I felt after this week was that the team wasn't too busy explaining, working on the bottom line -- refining the architecture, refining the efficiency, putting things together the same way.
This time, the attitude of literacy is brilliant: I'm chasing, chasing very seriously, things are here, randomly measured.
And they said unto him, Thou shalt not be lifted up above the clouds; thou shalt tread upon the mountains by the snow.
It's too early to conclude that the opponent on the list is stronger than the other. But in the snow, this series of footprints, one step at a time, can't be fake.
Currently, the Windows 5.5 Flash Preview is available on the Baidu Intelligent Cloud Sail platform, which developers can call through the API. The price is 2 yuan per million tokens for entry, 1 yuan per hour, 8 yuan per hour, 4 yuan per hour, and the current limit is five times.
Don't just listen to me. Let it play you a game.
Address of the experience:
