11-Hour Marathon Yields UE5 Masterpiece! Wenxin 5.5 Flash Preview Released

Crammed 11 Hours to Crank Out a UE5 Masterpiece! Wenxin 5.5 Flash Preview Released

  Xinzhiyuan Report 

Today, Baidu Wenxin unveiled its latest model:Wenxin 5.5 Flash Preview (ERNIE 5.5 Flash Preview), officially released!

This model: fast, cheap, open-source.

Fast, really fast. Over 200 tokens per second, 280 billion parameters, using only 14 billion per request.

Cheap, truly cheap. API costs per million tokens: input ¥2 during peak, ¥1 off‑peak; output ¥8 during peak, ¥4 off‑peak, currently limited to a 50% discount.

Today we’re launching Baidu Cloud Qianfan first, with direct API access; the model will be released to the global open‑source community on November 25.

Wenxin 5.5 Flash Preview receives a comprehensive upgrade in programming, 3D modeling, knowledge work, and content creation.

In multiple benchmarks such as Terminal‑Bench and AutomationBench, it delivers competitive scores:

Testing the truly hands‑on Terminal‑Bench 4.0 in the terminal, 46.5 points; testing the cross‑application AutomationBench, 61.9 points; testing the real Excel work with SpreadsheetBench V2, 36.4 points.

The * in the figure indicates Baidu’s self-test scores.

There are still projects being pursued, but don’t forget its positioning: a Flash model that emphasizes speed and efficiency.

2.8 trillion parameters, activating only 140 billion at a time; 1 million‑token context; can generate over 200 tokens per second.

With this positioning, it stands shoulder‑to‑shoulder with big names like Claude Opus 5.5 and GPT‑6.1 Sol, and even snatches two first‑place spots.

I’ve always taken benchmark scores with a grain of salt. The other half depends on how it performs in real work.

First, the thing that really surprised me: “Yun Xiu Xing”.

A third‑person Eastern martial‑arts game. Mist‑shrouded stone peaks of Zhangjiajie, an ancient temple atop a cliff. The game engine is Unreal Engine 5.8, the underlying code is C++, and it was fully completed by Wenxin 5.5 Flash Preview in a single 11‑hour task.

Eleven hours isn’t surprising. What’s surprising is that Wenxin 5.5 Flash Preview is a pure‑text model and can’t display any visuals.

The most common mistakes in making 3D games are characters floating in midair, running with twisted direction, camera positions misaligned—issues that anyone can tell are wrong at a glance.

Its solution is straightforward and clever:If you can’t see it, then measure it.

For example, if a character is hanging in mid‑air, you don’t need to look at the screen—just write a small program to measure how high the feet are above the ground and you’ll know. It turns the visual issues that pop up during playtesting into testable code that can run and produce data, then tests, fixes, and retests them repeatedly until the game is fully built.

Along the way it also did a dirty job: merging assets that span several generations of the Unreal Engine into a single UE 5.8 project. Different engine generations often have incompatible formats and interfaces. Anyone who’s made games knows this task makes you want to quit.

Honestly, what impressed me about this case wasn’t the visuals, but the workflow.

Anyone who has written code knows: a newbie reports a bug with “something seems off,” while a veteran says “the return value on line 37 is off by 0.3.” Turning vague impressions into verifiable numbers is a skill only a master possesses.

From a Pelican to a W16:

Code is its master key

Working continuously for 11 hours without crashing, where does this stamina come from?

According to the official disclosure, it used over 14 million “truly runnable” code tasks during training, plus execution traces from various agent frameworks.

What you practice is the whole cycle of “write, run, see the error, then fix it”; just memorizing code is useless.

With a massive 1‑million‑token context window, everything you’ve read, tools you’ve used, and results you’ve obtained over dozens of hours stays in your mind. The major game engines—Unreal, Unity, and Godot—have all been specifically adapted.

How’s the practice going? We even came up with a challenge ourselves: create a third‑person 3D sword‑flying game, “Sword Soaring Through the Clouds.”

The game starts on a stone platform. The protagonist wears a blue robe, a light‑colored ribbon hangs from his back, his hair is tied in a bun, and he stands steadily on a flying sword—no need to cobble together a few blocks.

Looking ahead, warm golden clouds hold up a series of teal stone peaks, a stone arch spans across the midway, and ten golden halos line up all the way to the cliff‑top temple.

The key is that it’s actually doable. Fly through the 3rd and 7th rings, and the checkpoints are recorded automatically; the ring count, timer, and energy bar in the upper left keep running, and there’s a mini‑map in the lower right. Finally you land in front of the temple with red pillars and double eaves, and the results panel pops up: all ten rings cleared.

Coincidentally, we also had it shoot an educational video. The title is just one sentence: What would the world be like if the Moon were only half its current distance from Earth?

The final cut is a little over a minute, vertical format. It opens with a massive Moon pressing against the sea surface, and the line “The first thing to go out of control isn’t the night sky, it’s the sea” pulls the viewer in.

What surprised me was its sense of proportion. When discussing tides, it deliberately added a sentence: the tidal force is 8 times, not that the tide height is necessarily 8 times. No one asked, it calculated the Roche limit itself and told you the Moon won’t be torn apart, still more than ten times away from that line.

I ran a few key numbers through a calculator, and they all matched. A Flash model does the work of an entire science‑communication video team: someone researches, someone does the math, someone creates graphics, someone edits the footage.

The same skill set works just as well in a different test.

There’s a widely circulated cheesy test in the developer community: ask the model to code‑draw a pelican riding a bicycle. Whether the drawing looks decent basically reveals the model’s capability.

Others submit a single image. Wenxin 5.5 Flash Preview submits a whole video.

“Pelican Test”, a 117‑second 2D animation written entirely from scratch in Python. A white pelican rides a bicycle, starting from the Sahara rock art of 4000 BCE, pedaling through Han dynasty stone reliefs, Javanese shadow puppetry, 8‑bit pixels, switching through 20 artistic styles, riding straight into the code editor of the large‑model era, and finally arriving at the “21st Pelican Launch”. By the time it reaches Van Gogh’s “Starry Night”, the swirling sky twists in sync with its pedal strokes.

The soundtrack uses 22 recordings, ranging from bronze drums and guqin to the Mario theme. No editing was done manually: the model generated the editing script itself, aligning each transition point scene by scene. (Concept design by B‑station creator @近步抱告_, authorized.)

The problem setter just wanted a pelican, and it delivered a brief history of pelicans. That’s a bit unfair.

The video editing is professionally practiced.

According to official disclosure, training focused on scenarios such as popular science, comic dramas, and creative editing, building over 150,000 creative environments. From concept, script, storyboard to voice‑over, it’s a full‑chain process, producing about 5,300 minutes of footage in total.

Switch to professional productivity software. This tests spatial awareness: given only a textual description, it must infer the shape, scale, and positional relationships on its own.

In the 3D modeling software Blender, two Wenxin agents start working separately—one builds an airport, the other constructs an aircraft.

If you’re building an airport, go to a map platform to get layout data, download supporting assets such as vegetation, vehicles, and ground‑service equipment; if you’re building an aircraft, assemble the fuselage, wings, cockpit, and double‑deck cabin piece by piece. Finally align the scale and camera on both sides and render the shot. It’s just like a small outsourcing team, without any hassle.

In the engineering design software FreeCAD, a W16 engine was built using a Python script; when the crankshaft turns, the pistons in the four rows of cylinders move up and down.

It also modeled editable 3D models of the east and west towers of Baidu International Building in Shenzhen, down to interior layouts and MEP pipelines; and drew inspiration from the Yutu‑2 rover to design a lunar concept car with a swaying suspension and rotating solar wings.

Seeing this, I finally realized.

Games, animation, aircraft, engines—seem unrelated, but underneath it’s the same thing: writing code.

Modeling is done by it writing scripts to command Blender; editing clips is done by it writing programs to schedule audio‑video assets. Wherever code is reachable, it can get its hands on it.

So Baidu lists programming ability as the top highlight of this upgrade, and I’m convinced. This isn’t just a skill; it’s a master key.

This key can even open the office door. The first two items at the beginning are testing exactly this kind of task.

It independently wrote a 24‑page autonomous driving research report, digested 4,169 U.S. federal accident reports, and dissected the safety data of Waymo and Tesla’s two approaches. It also created a personal wealth‑planning tool that simulates 1,000 random market paths to calculate whether your retirement fund will be sufficient.

This is also trained. For knowledge work such as finance and law, the official has set up 300,000 workspaces, prepared 10 million reference documents, covering more than 40 file types.

It even composed its own arrangement, producing a 6‑minute‑35‑second symphony.

This “quick”

Part of it was cut out by the AI itself

Being capable is one thing; being fast and affordable is another.

An AI agent’s tasks often involve hundreds of steps; if each step is a bit slower or more expensive, the end‑of‑month bill can give you a heart attack.

Wenxin 5.5 Flash Preview is architecturally fixated on the word “efficiency”:

Reading long texts saves VRAM: The model has 43 layers in total; 37 layers only read the “summary” when processing historical content, compressing every 128 tokens into one entry. The remaining 6 layers read the original text word by word, focusing only on the most critical positions. It’s like a project team where most members read meeting minutes, and only a few check the original documents. Even with a million tokens fed in, the VRAM can still handle it.

Select experts to avoid mistakes: Each layer has 512 “experts”, only 10 are activated each time, plus one permanent shared expert; the 280 billion parameters are thus compressed to use only 14 billion per inference. Task assignment also matters: previously a single “scheduler” (router) decided who to activate, leading to oversights—unselected experts had no chance to correct errors, making mistakes hard to fix; now multiple schedulers score together, preventing any single authority from dominating.

Draft while saving time: There’s a clever design called ECHO: it doesn’t add extra modules, it directly uses the model’s own final layer, and when writing the next token, it conveniently drafts the next few tokens as well; if they check out, just use them.

The architecture diagram also includes a few awkwardly named components: the mHC widening layer’s information channel between layers, the MultiMax module for the final word‑selection step, and the R2L tokenizer that splits numbers into groups of three digits from right to left, making calculation easier.

Only when several components are combined do we achieve a generation speed of over 200 tokens per second. This Flash of Wenxin is truly deserving of its name.

The more interesting part comes later.

On November 9, Baidu’s Wenxin disclosed progress in the RSI (Recursive Self‑Improvement) direction: an automated research system called ERNIE Evolve. In plain terms, it lets AI act as a researcher, generating ideas, modifying code, running experiments, and the results still have to pass independent evaluation—no cheating.

The practical task it received was to accelerate inference for Wenxin 5.5 Flash. The submitted work looks like this:

In the low‑latency scenario of the domestic Kunlun P800 chip, generation speed increased to 623.6% of the baseline. In the “deterministic inference” used for reinforcement‑learning training (the same request must produce identical tokens and probabilities each time, at the cost of slower speed), throughput rose by 33.7%.

The most skillful part is the ultimate decoding. In this scenario, human engineers have already specially optimized it, the easy gains have long been exhausted, and the remaining single-item improvements are often only 0.3% to 0.5%, even smaller than measurement noise.

It stubbornly tried seven directions and squeezed out a 14.9% speedup.

AI gets hands‑on to polish the next‑generation AI. This sentence used to live in sci‑fi novels, now it’s written into the development log of a domestic large model.

Three posts in seven days: No shame in catching up

Flip the calendar back; Baidu’s pace over the past seven days is surprisingly intense.

On November 5, Wenxin 5.5 Flash Base was open‑sourced, marking Baidu’s release of a trillion‑parameter‑scale model weight after a 16‑month gap. On November 9, the RSI automatic research progress was disclosed. On November 11, Wenxin 5.5 Flash Preview was officially launched. Three releases in seven days.

The next release is already scheduled: on November 25, the weights of Wenxin 5.5 Flash Preview will be open‑sourced.

And tomorrow, Baidu World 2026 will open in Beijing, with the theme “The Next Act, Intelligent Organization”. On the eve of the conference, they unveiled a new model, and the real show is probably still to come.

Let’s be realistic.

In the past two years, there have been many voices about Wenxin, with both expectations and doubts.

After going through everything this week, my impression is: this team isn’t busy explaining, they’re heads down getting things done. Refining the architecture, squeezing efficiency, presenting each piece of work. Even a few tenths of a percent speed‑up, someone is using AI to fine‑tune it step by step.

Chasing isn’t shameful. This time Wenxin’s stance is quite open: I’m chasing, and I’m doing it seriously. The results are here, feel free to test.

There’s a saying I’d like to give it: No one will lift your soaring ambition, you climb the snowy peak yourself.

Whether you’ve reached the summit yet is still too early to conclude; the competitors on the leaderboard are each stronger than the last. But these footprints in the snow were made step by step, no fake. This praise, Wenxin can live up to.

Currently, Ernie 5.5 Flash Preview is available on Baidu Intelligent Cloud Qianfan platform, and developers can access it via API. Pricing is 2 CNY per million input tokens during peak hours, 1 CNY off‑peak; 8 CNY per million output tokens during peak, 4 CNY off‑peak, currently a limited‑time 50% discount.

Don’t just listen to me. Go have it whip up a game for you.

Experience URL: https://ernie.baidu.com/web/new-chat

Edited by: Solomon

Instant ASI

⭐Like, share, and view with one click⭐

Turn on the star to get instant updates from Xinzhiyuan!