The stock price surged by 32%, with GLM-5 topping the global open-source charts. A 25-minute continuous shot showcases the complete system.

On the late night of February 7th, a mysterious model codenamed “Pony Alpha” quietly went live.

Soon after, it caused an explosion online.

Input a piece of “mountain of messy code” that took a day to modify, and it effortlessly reconstructed the architecture; input a simple prompt, and it outputs a complete web app with 35 radio stations and a silky UI.

This extreme engineering capability directly confirmed Andrej Karpathy’s assertion from a few days ago:

Vibe Coding is a thing of the past. The new game rule has only one name—

Agentic Engineering.

Immediately afterward, Opus 4.6 and GPT-5.3-Codex “crashed” into each other late at night the next day, both focusing solely on “long-term tasks and system engineering.”

Just when everyone thought this was another closed-source giant’s solo act, the mystery of Pony Alpha was unveiled—

It is GLM-5.

The world’s first open-source model to hit this track, directly competing head-to-head with Silicon Valley giants in system-level engineering capabilities.

After the reveal, the stock price of Zhizhi surged by 32%!

The first in the world to open source! The “Opus Moment” for domestic models

After truly hands-on experience, our only feeling is: truly powerful!

If Claude Opus represents the pinnacle of closed-source models, then the release of GLM-5 undoubtedly marks the arrival of China’s own “Opus Moment” for open-source models.

On the day of release, more than ten games and tools developed by developers based on GLM-5 were simultaneously showcased and available for testing. These applications will also be gradually launched on major app stores.

This means GLM-5 is transforming “AI programming” into “AI delivery,” seamlessly bridging productivity tools and commercial products.

For example, this project called “Pookie World.”

It’s a digital parallel universe driven by GLM-5, endowed with autonomous intelligent agents with real narrative integrity and life motivation through multi-layered bio-psychological frameworks.

There’s also a replica of “Minecraft,” with identical effects and gameplay.

We also used Claude Code as a shell, directly connecting to GLM-5’s API for multi-dimensional testing.

Whether it’s a full-stack Next.js project or a native MacOS/iOS app, it can handle the entire process from requirements analysis, architecture design, to coding and end-to-end debugging.

Having done many projects, there’s an inexplicable feeling:

To some extent, GLM-5 might be a model capable of changing the industry landscape.

· Challenging complex logic: “Infinite Knowledge Universe”

If you think building a webpage is easy, try asking AI to handle an “infinite stream” project with strict JSON format requirements and dynamic rendering.

Take our first test project, “Infinite Knowledge Universe.”

It’s a typical complex front-end and back-end separated project, involving React Flow dynamic rendering, Next.js API routing, and very strict JSON output requirements.

GLM-5’s performance here is stunning.

It not only completed the entire project file structure in one go, but also surprised us with its debugging logic.

When encountering a rendering bug, we only said, “The page is still black, and the first content didn’t appear during initialization…”

GLM-5 immediately pinpointed it as a load timing issue and quickly provided a fix.

The full prompt was as follows:

Infinite Stream·Concept Visualization

Core concept: This is an “endless” mind map that never stops. Users input any keyword (like “quantum physics” or “Dream of the Red Chamber”), and the system generates a central node. Clicking any node, AI expands its child nodes in real-time.

Stunning moment: Users feel like they are interacting with an omniscient brain. When they casually click on a rare concept, and AI can still accurately expand the next level, the sense of “infinite exploration” is very shocking.

Visual and dissemination:

• Use React Flow or ECharts to create dynamic, draggable node networks.
• Color schemes in Cyberpunk or minimalist styles, perfect for screenshots and sharing on social media.

Feasibility plan:

• Frontend: React + React Flow (for drawing).
• Backend: Next.js API routes.
• Prompt strategy: No need for complex context memory, just let AI generate 5-6 related sub-nodes for the “current node” and return in JSON format.
• Key challenge: Ensuring the model outputs stable JSON format (an excellent scenario to test the model’s instruction-following ability).

· Building a more complex middle-platform project in 11 minutes

Next, increase the difficulty: ask it to develop a project called “Soul Mirror,” a psychological analysis app.

The requirements are divided into two steps:

Step 1: Logic design—play the role of a Jungian psychology expert, outputting a JSON containing analysis text and visual parameters.

Step 2: Frontend implementation—generate SVG dynamically based on parameters, creating tarot card-style visuals.

Prompt

Step 1: Logic Design

We want to develop a psychological analysis app called “Soul Mirror.”

Interaction flow:

1. Welcome page: user inputs current state or confusion.

2. Analysis page: AI asks 2 deep follow-up questions to guide user exploration.

3. Result page: based on the dialogue, AI generates a “soul card.”

Please design the core prompt (System Instruction): ask the model to act as a Jungian psychology expert. In the final step, the model needs to output a JSON containing:

• analysis: psychological analysis text.

• visualParams: a set of parameters for generating abstract art images (e.g., colorPalette (hex color array), shapes (circle/triangle/wave), chaosLevel (numeric)).

Step 2: Frontend implementation and SVG rendering

Write Next.js frontend code. Focus on implementing a ResultCard component.

Requirements:

1. Accept visualParams from Step 1.

2. Use SVG to draw graphics dynamically. For example: if chaosLevel is high, use irregular paths; if colorPalette is warm, use gradient orange-red backgrounds.

3. The card layout should be exquisite, like a tarot card: center is a dynamic SVG pattern, bottom shows user’s name and AI’s “soul motto.”

4. Add a “Save as Image” button (using html-to-image library).

Throughout the process, its understanding often makes people doubt whether they are using Opus 4.5.

But a quick glance confirms—it’s indeed GLM-5.

· 25-minute one-shot “Agentic Coding”

To further test GLM-5’s capabilities, we asked it to simulate a real user without using APIs, to build a monitoring system for platform X.

Result: 25 minutes, one continuous shot.

As seen, GLM-5 autonomously calls various tool agents during operation, plans tasks, breaks down steps, and even checks documentation to fix errors.

This long-term logical coherence is something previous open-source models couldn’t even dream of.

· From image to app, truly impressive fidelity

Finally, we took a screenshot of an open-source project by the father of OpenClaw (a tool for AI quota statistics), and directly fed it to GLM-5:

Make me a MacOS app based on this.

In no time, it “recreated” a similar product.

Though the data was mocked, the UI layout and interaction logic were almost perfectly replicated.

This demonstrates not only visual understanding but also the engineering ability to convert visuals into SwiftUI code.

Master’s manual: Recreating “Ghetto Cursor” in 1 day

To test the engineering limits of GLM-5, a senior developer decided to go big:

From scratch, build a desktop UI AI programming assistant—GLMLIFE.

It’s essentially a simplified version of Cursor.

The task was handed to GLM-5, which didn’t just start coding wildly but first produced a professional architecture document (PLAN.md), with very mature technical choices:

Using a monorepo structure, splitting the project into three core packages:

• Core: handles agent engine and LLM adaptation;
• CLI: manages command-line interactions;
• Desktop: based on Electron + React 18 for the desktop app.

From Zustand state management to Tailwind styling, and complex IPC communication, GLM-5 acts like a seasoned CTO, making clear technical decisions.

Originally expecting three days to set up the environment, it only took one day to get from environment setup, core logic, to Electron packaging.

Open GLMLIFE, and it’s hard to believe this is AI “architecting” in just one day.

Why can it become “Open Source Opus”?

Globally, Claude Opus 4.6 and GPT-5.3-Codex are highly sought after because of their strong “architecture” capabilities.

• Opus 4.6’s brutal aesthetics: 16 AI clones work independently, taking two weeks to build a Rust compiler with 100,000 lines of code, passing 99% of GCC stress tests.
• GPT-5.3’s self-creation: It’s OpenAI’s first model to “participate in its own creation,” involved in training and deployment before “birth.”

But all this has a fatal premise: they are both closed-source and expensive.

Now, with the release of GLM-5, China’s open-source large models have violently broken into the Agentic era.

It directly targets the least willing domain of closed giants—system-level engineering—and launches a “substitute” attack.

1. The emerging “Backend Architect”

Zhizhi team knows very well that the open-source community isn’t short of models that can write Python scripts; what’s missing are models capable of handling dirty, tedious, and big tasks.

GLM-5 has significantly enhanced weights for backend architecture design, complex algorithm implementation, and stubborn bug fixing during training, and also features a strong self-reflection mechanism.

When compilation fails, it acts like a seasoned engineer, analyzing logs, pinpointing root causes, modifying code, and recompiling until the system runs smoothly.

2. If it’s doing the work, it should count the cost

Matching Opus in performance, with open weights, makes GLM-5 somewhat shake the walls built by Anthropic and OpenAI.

• Local deployment: It can run in fully isolated intranet environments and can be fine-tuned for private frameworks, becoming the most knowledgeable in its own codebase.
• Cost control: Users can run a powerful coding agent on consumer-grade GPU clusters, no longer worrying about costs with each test.

Top of the SOTA

This evolution of GLM-5 can only be described with two words: violence.

Since it’s a foundational model for complex system engineering, scale must be maximized.

Parameter count jumped from 355B (activation 32B) to 744B (activation 40B), pretraining data from 23T to 28.5T.

Besides “big,” it also needs to be “cost-efficient.”

It’s well known that the most expensive part of running an agent is tokens.

To address this, GLM-5 integrated DeepSeek Sparse Attention for the first time.

This allows it to handle ultra-long contexts with “lossless” memory and significantly reduces deployment costs.

Another “black tech” is the new asynchronous reinforcement learning framework Slime.

Coupled with large-scale reinforcement learning, it turns the model from a one-time tool into a “long-distance runner” that gets smarter over time.

As for benchmarks, it’s all hardcore:

• Code capability
  SWE-bench verified score hits 77.8, Terminal Bench 2.0 scores 56.2, both first in open source. This surpasses Gemini 3.0 Pro and even closely rivals Claude Opus 4.5.

• Agent capability
  BrowseComp (web retrieval), MCP-Atlas (tool invocation), and τ²-Bench (complex planning) all top open-source rankings.

The most interesting is Vending Bench 2 (automatic vending machine management test).

In this test, the model must run a vending machine business for a year entirely on its own.

Guess what? GLM-5 earned $4,432 by the end of the year—almost catching up with Opus 4.5.

In internal Claude Code evaluation sets most developers care about, GLM-5 significantly outperforms the previous GLM-4.7 (average improvement over 20%).

Real-world experience already approaches Opus 4.5.

Building AI with AI

Of course, GLM-5’s ambition isn’t just about models but also about reconstructing our programming tools.

The globally popular OpenClaw showed the potential of AI to operate computers.

This time, Zhizhi launched AutoGLM version of OpenClaw.

Using the original version requires days of environment setup; now, it’s deployable with one click on the official website.

Want a “digital intern” that monitors Twitter, organizes news, and even writes scripts 24/7? Just a click away.

Also released is Z Code—

A new generation of development tools born entirely from GLM-5’s capabilities.

In Z Code, you just specify your needs, and the model automatically decomposes tasks, even launching multiple agents to work concurrently: coding, running commands, debugging, previewing, and even Git commits.

You can even remotely command desktop agents via your phone.

It’s worth noting that just as OpenAI used Codex to create Codex, Z Code itself was also developed with full participation of the GLM model.

Victory of domestic computing power

Behind GLM-5’s global traffic surge and rising agent demand, a group of “unsung heroes” silently supports the massive computational load.

To ensure every line of code and every agent plan is stable, GLM-5 has deeply integrated with domestic chips like Huawei Ascend, Moore Threads, Cambrian, Kunlun, Muxi, Suiyuan, and Hygon.

Through fine-grained optimization at the operator level, GLM-5 can run with “high throughput and low latency” on domestic chip clusters.

This means we not only have top-tier models but are no longer “necked” by technology.

Conclusion

Spring 2026, large programming models finally shed their childishness.

Karpathy’s so-called “Agentic Engineering” essentially imposes a more rigorous “interview” on AI:

• Previously (Vibe Coding): As long as you can write pretty HTML, I’ll hire you.
• Now (Agentic Coding): You need to understand Linux kernel, microservice call relationships, how to refactor code without crashing online, and also plan tasks and fix bugs yourself.

GLM-5 isn’t perfect.

But in the core challenge of “building complex systems,” it is currently the only open-source player capable of riding this wave of “Agentic tide.”

Vibe Coding is over.

Stop asking AI “Can you help me write a webpage?” That was 2025’s story.

Now, ask it: “Can you help me refactor the core module of this high-concurrency system?”

GLM-5, Ready to Build!

Easter Egg

GLM-5 has been included in the Max user package, with Pro support coming within 5 days!

And Zhizhi just announced a price increase—tokens are definitely going up this year!

Source: Xin Zhiyuan

Risk warning and disclaimer

Market risks are present; investments should be cautious. This article does not constitute personal investment advice and does not consider individual user’s specific investment goals, financial situations, or needs. Users should consider whether any opinions, viewpoints, or conclusions herein are suitable for their particular circumstances. Invest accordingly at your own risk.