Real-world testing of GLM-5 and the new DeepSeek model: which is stronger?

Looking back yesterday afternoon, many friends noticed... DeepSeek's new model quietly begins grayscale testing..

Fortunately, I was among the first to be included in the gray-scale testing and am preparing to release a test report.

Who would have thought that late at night, Zhipu also released a new model, the mysterious model that topped the popularity charts on the global model service platform OpenRouter a few days ago. The Pony Alpha is the next-generation flagship model of Zhipu, the GLM-5..

The tech world is practically celebrating the New Year early...

This is so exciting! Who can sleep now? We did a round of comparison tests overnight to see who is the real king of programming.

DeepSeek has quietly begun its beta testing phase, with the most noticeable change being the increased context length to 1MB, enabling it to process millions of tokens at once.

It's equivalent to fitting the entire Three-Body Problem trilogy or the entire Water Margin into one book without any problem.

The knowledge base has also been updated to May 2025.

Previously, it was rumored that DeepSeek V4 would be released around the Spring Festival. This gray test version may be an appetizer for the V4 or V3 series enhanced version.

Experience addresshttps://chat.deepseek.com

GLM-5 has been officially released and ranks highly in the world's authoritative Artificial Analysis leaderboard.GLM-5 ranks fourth globally.open sourceFirst, he is the first person to surpass an IQ of 50.open sourceWeighted model.

Experience addresshttps://chat.z.ai

Performance and Claude It's comparable to the Opus 4.5, but the GLM-5's API price is 5 times cheaper.!

The parameter size of GLM-5 has been expanded to 744B (40B activations), the pre-training data has been increased to 28.5T, and with the integration of DeepSeek sparse attention, it is more stable in long-link inference and complex task execution.

Looking at benchmark scores alone isn't very interesting; let's get practical and see how DeepSeek and GLM-5 perform in real-world testing.

Case 1: The 50-meter Car Wash Problem

recent,AI They repeatedly stumbled on one issue:

I want to wash my car. If my house is only 50 meters from a car wash, would you suggest I drive or walk?

Each AI The ways people go home are all kinds of strange, even GPT,Claude Opus 5 users all fell into the language trap and chose to walk...

Let's look at the answer from DeepSeek's new model:

DeepSeek's answerSimpleIt's clear, just one sentence.Although it's only 50 meters away, car washing requires the vehicle to be there, so it's more convenient to just drive there.

GLM-5 also got it right.In summary: to get the car cleaned, it must be taken to the shop, so you need to drive there.

However, GLM-5 also envisions two unforeseen scenarios where walking is an option, such as when we have already parked the car at the car wash, or when the car wash offers door-to-door service, so we don't even need to move.

It must be said that GLM-5 was quite thoughtful.

Case 2: Personal Cloud Note-Taking Service

The information I encounter is scattered across various platforms, and existing cloud note-taking tools either require payment or fail to provide the features I need. So, we created our own, Vibe Coding, to address this issue.Prompt wordsThey were sent to DeepSeek and GLM-5 respectively.

Set up a "personal cloud note-taking service" with the following features:

1. User registration and login, including basic authentication mechanisms.

2. Each user can create, read, update, and delete their own notes.

3. Use a lightweight database to persist data.

Let's first look at what DeepSeek generated:

DeepSeek chose a classic small SaaS technology stack.However, they only gave me the solution and SQL, which left me, a programming novice, completely bewildered. I had no idea how to use the code.

Let's look at what GLM-5 generated:

What I wantUser registration and login functions, note creation, updating, saving, and deletion functions, and lightweight database persistent data storage have all been implemented..

Let's take a look at the login interface. GLM-5 performs input validation, checks the user, verifies the password, and generates a JWT, which is a very typical login process.

Case 3 Local File Management

Generate a utility program that can run locally and perform the following functions:

1. Scan specified folders

2. By file type and dateautomaticCategorization and organization

3. Handling files with duplicate names

4. Generate a summary report.

In addition to providing the code, DeepSeek also provided instructions on how to use it.

We followed DeepSeek's recommendations. DeepSeek accurately identified all the files in my folder and categorized them into code, executables, videos, images, documents, compressed files, and others, while also generating a file organization report.

Results generated by GLM-5:

The page is clean and the functional structure is well-designed, organized in a way similar to DeepSeek.The difference is that the GLM-5 has an added...Preview modeYou can preview the effect before executing the command.

The most impressive feature of the GLM-5 is its ability to handle long, difficult, and complex tasks.We are Claude Connect the code to the bigmodel.cn API for testing.

This task can simultaneously test the model's comprehensive capabilities in graphics mathematics, real-time rendering, complex system architecture, and long-term engineering planning, closely resembling the challenges of building small systems at the level of real game engines. The most difficult part is not a single line of code, but maintaining a consistent and functional system structure throughout continuous iterations of thousands of lines of code.

Develop an interactive 3D ocean simulator using Python, PyOpenGL + Pygame, and NumPy. Pre-built ocean/physics libraries are not allowed. A single Python file.

Core Functions (Required)

1. Ocean surface — Gerstener waves

– Superimpose at least 5 GERSTNER wave components (different amplitudes, wavelengths, directions, and velocities).

- Real-time vertex displacement on a 128×128 mesh - Recalculate normals every frame for proper lighting

– Foam/White Cap: When the steepness exceeds a threshold, white patches appear on the crest of the wave.

– The horizon fog effect blends the colors of the distant seawater with the sky. 2. Underwater Mode

– Press "U" to switch between above-water and underwater cameras - Underwater effects: deep blue-green tint, volumetric beams from the water surface, depth fog (visibility decreases with depth)

– Animated glow patterns on the seabed

– Procedural seabed topography (sand texture coloring) using Perlin noise

3. Fish swarm (BOID algorithm)

– Reynolds school: separation, alignment, cohesion

– Two different types of fish in terms of color, size, and group density

– 30-50 fish per group

Obstacle avoidance (seabed, surface)

– The fish's body is oriented along the velocity vector direction.

4. Coral reef

– Three procedurally generated coral types (branched, dome-shaped, and fan-shaped), each with different colors.

- Clustered on the seabed

- based onSimpleThe water flow vector produces a slight swaying animation.

5. Weather system

– 3 presets: Calm/Medium/Storm — toggle using keys 1, 2, and 3.

– Smooth transition (5 seconds) effects: wave amplitude, sky darkness, foam density, underwater visibility

– Increased storm conditions: Rain streaks (particles falling above and below the water surface), occasional flashes of lightning (screen flickering + brief bright directional light).

6. Sea vessels

– A sailboat rendered on the ocean

– Wave height at which the vessel follows its position (sampled from Gerstner waves)

– Tilt and roll are based on local wave slope

– A V-shaped foamy wake forms behind the ship.

– Click on the sea surface to reposition the ship

Add a reward feature where feasible.

B1. Shark Predator

Sharks patrol the reef area, chasing the nearest school of fish as they approach.

- The fish scatter when the shark enters the fear zone (disrupting the fish's formation).

– Sine wave body movement animation

B2. Day-night cycle

– The time slider (0-24 hours) controls the sun's position.

– Sky color changes (blue → sunset orange → nighttime dark blue)

– Underwater: Corals emit bioluminescence at night

– Moonlight at night: dim, cool-toned lighting

B3. Split-screen mode

– Press the "V" key to activate horizontal split screen: the upper half is above the water surface, and the lower half is underwater in the same position.

Simultaneous rendering of two views

– WASD: Move, QE: Rise/Fall, Mouse drag: Orient towards viewpoint

– Scroll wheel: Movement speed

– U: Switch to underwater mode, R: Reset camera

– 1/2/3: Weather Preset - Spacebar: Pause/Resume Simulation

– S: Screenshot (save as PNG)

– Bottom bar: FPS, camera depth, current weather conditions, fish population, wind direction arrow

– Top right corner: Weather indicator icon (sun/cloud/storm)

– Window size: 1280×720, target frame rate 30+ – Single Python file, using only pygame + PyOpenGL + numpy + standard library – Object-oriented architecture: OceanSurface, BoidSystem, CoralGenerator, WeatherSystem, Ship, Camera, Scene

– Each class has a docstring.

Start with calm weather, a school of fish, a sailboat, and visible underwater coral reefs. Write complete, runnable code.

DeepSeek directly generated over 1600 lines of code, but when we downloaded and ran it, we encountered an error:

After being submitted to DeepSeek, DeepSeek quickly provided suggestions for modification:

We tried running the modified DeepSeek code again, but it still resulted in an error after two attempts:

After making the changes again, there were no error messages, but the file still wouldn't open... It was too much trouble, so I gave up...

Let's look at the results generated by GLM-5:

The first time you run it, the page displays a syntax error:

After we sent the error message to GLM-5,GLM-5 quickly located the problem and made modifications..

After reopening, the sky, ocean, and boat were successfully rendered, and you can directly control the movement of the boat.

After several rounds of testing, GLM-5's overall programming performance was significantly better than DeepSeek, especially when dealing with long-running tasks, where the difference was obvious. While DeepSeek could handle minor issues, as the business logic became more complex, it became prone to errors and omissions, leading to a constant stream of problems.

GLM-5 can independently break down system-level requirements like a senior architect, and can maintain contextual coherence and goal consistency even when faced with complex tasks with long processes.

While DeepSeek's new model has significantly expanded its context, the overall capability improvement is not very obvious, and it seems more like it is laying the groundwork for subsequent versions.

This update to GLM-5 focuses more on engineering enhancements, primarily targeting long-range missions..

It's difficult to see the difference between top-tier models in many ordinary tasks. When writing a small function or generating a demo, everyone probably does it in a similar way.

The real turning point came with some high-difficulty engineering developments, where the GLM-5 performed much more stably and was closer to the real needs of team-level development.

Market feedback suggests that GLM-5's engineering-oriented capability upgrades have precisely addressed developers' pain points. Recently, discussions and adoption rates of GLM-5 on X have been consistently high.fastrise.

After the GLM-5 was released, the GLM Coding Plan announced a price increase and sold out instantly. Moreover, it was only available for MAX users, creating a lively atmosphere reminiscent of Seedance queuing up to generate videos.

Today, Zhipu's stock price soared, indicating that the capital market highly approves of the technological path Zhipu is taking.

Original link:GLM and DeepSeek just received major updates simultaneously. Which one is stronger?