Ox Alpha Revealed: GLM-5.3-Flash Is Now Available on ApiHub

Remember Ox Alpha?

The mysterious stealth model that suddenly became popular among developers for coding, reasoning, and agent workflows?

We finally know what it is.

Ox Alpha was GLM-5.3-Flash.

Before the official release, Z.ai anonymously deployed the model as Ox Alpha for large-scale real-world testing.

Now the model has officially launched as GLM-5.3-Flash — and it is also available on ApiHub.

But the interesting part isn't just the name reveal.

GLM-5.3-Flash introduces a very different approach to building a powerful frontier model:

More intelligence, much less compute.

Let's take a closer look.

The First Native Multimodal Model in the GLM-5 Family

GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series.

It was trained from a new base model using a large-scale multimodal training corpus, rather than adding vision as a separate capability later.

That means visual understanding isn't just another input format.

It can become part of the model's reasoning and execution loop.

This is particularly important for:

• Coding agents
• Browser agents
• Computer-use agents
• Frontend development
• UI debugging
• Visual document processing
• Image understanding
• Professional productivity workflows

For developers, this opens up workflows where the model can not only generate something, but also look at the result and improve it.

320B Parameters — But Only 18B Active

GLM-5.3-Flash has:

• 320B total parameters
• 18B activated parameters

That's an interesting combination.

Instead of activating hundreds of billions of parameters for every token, only a relatively small portion of the model participates in each inference step.

Compared with earlier GLM architectures of a similar total size, Z.ai also reduced the number of layers significantly.

The goal is clear:

Keep frontier-level intelligence while making inference much more efficient.

And according to Z.ai, GLM-5.3-Flash already outperforms GLM-5.2 across multiple benchmarks and real-world tasks despite using a much more efficient architecture.

Sparse Attention + Linear Attention

One of the most interesting technical changes is the attention architecture.

GLM-5.3-Flash combines:

Sparse Attention + Linear Attention

instead of relying entirely on traditional attention mechanisms.

The two approaches play different roles.

Linear Attention

Linear attention focuses on efficiently modeling local dependencies.

Sparse Attention

Sparse attention uses a lightweight indexing mechanism to retrieve important information from the broader context.

Together, they allow the model to maintain strong long-context capabilities without paying the full computational cost of conventional attention.

The result is significant.

Compared with GLM-5.3, GLM-5.3-Flash reduces:

• Attention computation by 3.01×
• KV cache size by 4.44×

This becomes especially important for long-context applications.

Because as context windows grow, KV cache memory and attention computation quickly become major inference bottlenecks.

Why KV Cache Matters

For a simple chatbot conversation, KV cache optimization may not sound very exciting.

But think about an agent working with:

text
System instructions
        +
Large repository
        +
Documentation
        +
Tool results
        +
Browser state
        +
Screenshots
        +
Terminal output
        +
Previous actions
        +
Current task