GLM 5.3 Flash: Z.ai Ships Ox Alpha as a 320B Open Model

Z.ai released GLM 5.3 Flash on August 26, 2026, and it is not the model anyone was waiting for. For twelve days the expected next event from Z.ai was the open-weight release of GLM 5.3, promised “in about two weeks” at that model’s launch. What arrived instead was a different model entirely. GLM 5.3 Flash is a 320-billion-parameter mixture of experts that activates just 18 billion parameters per token. It is the first natively multimodal model in the GLM-5 series, and the first built on a new base since GLM 5.2.

It also arrived with an identity reveal. GLM 5.3 Flash is Ox Alpha, the anonymous model that had been quietly serving traffic on OpenRouter and drawing attention for its 1M-token context. Below is what the model actually is, what Z.ai’s benchmark chart shows once you include the tests it loses, and what it costs. It also covers why the weights that did ship are not the ones that were promised.

The Key Takeaways

• Released: August 26, 2026, with MIT-licensed weights on Hugging Face on day one.
• Size: 320B total, 18B active, built on a newly trained base rather than post-trained on GLM 5.2’s.
• A first for the line: natively multimodal, with a 1M-token context and a hybrid sparse and linear attention design.
• Price: $0.15 per million input tokens and $0.50 output, against $1.40 and $4.40 for GLM 5.3.
• Still missing: GLM 5.3’s own 744B weights. Z.ai’s Hugging Face organisation has no GLM 5.3 repository at all.

What GLM 5.3 Flash Actually Is

The name invites a wrong assumption. In most model families, “Flash” means a distilled or trimmed version of the flagship. That is not what this is. Z.ai’s own announcement laid out the specification in four lines.

A new base, not a post-train

This is the detail that separates GLM 5.3 Flash from everything Z.ai has shipped this year. GLM 5.3 took its entire gain from post-training on the same 744B base as its predecessor, with no new pretrain at all. GLM 5.3 Flash reverses that. Z.ai says the model “starts from a newly trained base model, with its architecture and training recipe redesigned around capability and efficiency,” trained on a fresh 30-trillion-token multimodal corpus.

So the two models sit on opposite sides of a design split. One is a bigger brain taught new habits. The other is a smaller brain built differently from the ground up. For context on how this family arrived here, our explainer on what GLM is and how Zhipu’s open line evolved covers the earlier generations.

What the hybrid attention buys

GLM 5.3 Flash is the first model in the GLM series to combine sparse and linear attention in one stack, which is where most of its cost advantage comes from. The published config makes the split visible: of 45 layers, 34 use linear attention and 11 use sparse attention, and the mixture of experts routes to 8 of 288 experts per token. Z.ai’s documentation puts numbers on it: against GLM 5.3, the design cuts attention computation by 3.01 times and KV cache size by 4.44 times. On a 1M-token context, the KV cache is usually what makes long sessions expensive to serve, so that second figure is the one that shows up on the invoice.

The model also adopts a technique called Manifold-Constrained Hyper-Connections, or mHC, which constrains how residual streams mix so that very wide connectivity does not destabilise training. Worth noting where it came from: mHC was published by a DeepSeek research team in late 2025. A Zhipu model shipping on a rival Chinese lab’s published method is a small but real illustration of how fast open research circulates between them.

GLM 5.3 Flash against GLM 5.3

Put side by side, these are not variants of one model. They disagree on almost every line.

Ox Alpha, Unmasked

Before the launch, an unnamed model called Ox Alpha had been running on OpenRouter with a 1M-token context, free during preview, and drawing notice for handling whole codebases without chunking. The reveal was confirmed twice on launch day. Z.ai’s announcement described the model as “previously previewed as Ox Alpha.” Zixuan Li, credited on the GLM-5 technical report, went further. Ox Alpha “was an early version of GLM-5.3-Flash,” he wrote, and the shipped release “delivers stronger performance and significantly better stability.”

The stealth slot is now empty. Ox Alpha no longer appears in OpenRouter’s live model catalogue, and z-ai/glm-5.3-flash was created there at 13:59 UTC on August 26, thirteen minutes before Z.ai posted the announcement.

Anonymous preview launches have become a standard tactic for getting real usage data before a name is attached, and they do not always end this tidily. We covered a messier example in the Hunter Alpha mystery model that fooled the industry.

One line in the announcement deserves care. Z.ai says the model was “running entirely on Chinese AI chips,” which describes how the preview was served. It is not the same claim as saying the model was trained on them, and Z.ai has not said that about this model. The distinction matters because Z.ai did train GLM-5 end to end on Huawei Ascend silicon, and the two facts blur together easily.

GLM 5.3 Flash Benchmarks

Every figure below comes from Z.ai’s own launch chart, including the scores it publishes for rival models. There is no independent evaluation of GLM 5.3 Flash yet. The one exception is GDPval-AA v2, which Z.ai’s footnotes say was run by Artificial Analysis rather than in house.

Where GLM 5.3 Flash leads

The cleanest result is against its own predecessor. GLM 5.3 Flash beats GLM 5.2 on all six tests, at less than half the size and roughly a tenth of the price. That is the specific claim Z.ai makes in the model card, and the chart supports it.

Against the closed frontier, two results stand out. AutomationBench at 48.8 puts it well ahead of Claude Opus 4.8 at 41.0 and GPT-5.6 Terra at 37.2. GDPval-AA v2, the one number here not produced by Z.ai, has it at 1773 Elo against 1582 for Opus 4.8, the highest score on the chart by a clear margin.

Where it still trails

The chart is less flattering than the launch-day summaries suggested, and the losses are worth stating plainly. On Terminal Bench 2.1 the model sits last but one at 84.3, behind Opus 4.8, Gemini 3.7 Flash and GPT-5.6 Terra. On Agents’ Last Exam it is last of the five plotted. On HLE with tools it trails both Opus 4.8 and DeepSeek. On AutomationBench, its strongest showing against Anthropic and OpenAI, Gemini 3.7 Flash still beats it at 52.3.

Read as a whole, GLM 5.3 Flash is not a frontier-beating model and Z.ai does not claim it is. The claim is that it gets close to Opus 4.8 on coding and agentic work while costing a fraction as much, and on these numbers that holds. For a broader view across vendors, see how the GLM line compares with Claude, GPT and Gemini.

GLM 5.3 Flash Pricing

This is where the model makes its case. Z.ai’s published rate card puts GLM 5.3 Flash roughly an order of magnitude below its own flagship, and a launch promotion currently halves even that.

All figures are per million tokens, from Z.ai’s developer pricing page. The 50 percent discount runs until midnight on September 9, 2026, Singapore time, and OpenRouter is already billing the promotional rate, which is a useful independent check that the discount is live rather than announced.

On the subscription side, GLM 5.3 Flash is included in the GLM Coding Plan with three times the quota of GLM 5.3, and off-peak calls, including all weekend, consume half the usual points. Our breakdown of GLM pricing and how the Coding Plan quota behaves covers the tiers in detail.

The Weights Z.ai Did Not Ship

On August 14, Z.ai launched GLM 5.3 without open weights and said they would follow in about two weeks, after a safety review of the model’s vulnerability-discovery capability. Two weeks from that date is now.

What shipped is a different model. As of this article, Z.ai’s Hugging Face organisation lists GLM-5.3-Flash and GLM-5.3-Flash-BF16, both created on August 25, and before those the most recent entries are GLM-5.2 and GLM-5.2-FP8 from June 16. There is still no GLM 5.3 repository.

That is not necessarily a broken promise. GLM 5.3 Flash is a different model with a different risk profile, so nothing about releasing it says the cyber review of GLM 5.3 has concluded. But the state of play has not moved. The strongest GLM model you can download is still not the strongest GLM model Z.ai has. Anyone whose plans depend on self-hosting frontier-class open weights should look at where the rest of the field stands in our guide to the best open-source AI models.

How to Use GLM 5.3 Flash

Access is unusually wide for a launch day. The model is on Z.ai’s own API as glm-5.3-flash, on the GLM Coding Plan, on OpenRouter, and downloadable from Hugging Face under MIT. Z.ai recommends running it with temperature at 1, top_p at 0.95 and reasoning effort at maximum, and notes that thinking mode cannot be switched off.

Image input works by adding an image_url content block to a message, and multiple images can go in one request. Video appears to be supported too, though less clearly documented. The published model config defines video tokens alongside image tokens, and OpenRouter lists video among the accepted input types. Z.ai’s own API guide describes only the image path.

Running it yourself

Z.ai publishes deployment recipes for vLLM, SGLang, TokenSpeed and KTransformers, so the serving side is well covered. The hardware side is where expectations need adjusting.

An 18-billion active parameter count does not make this an 18-billion parameter model to host. Mixture-of-experts routing picks a different subset for every token, so all 320 billion weights have to be resident and reachable. In practice that is a multi-GPU server, not a workstation, and certainly not a laptop. What the low activation count buys is throughput and serving cost once the model is loaded, not a smaller download.

Using GLM 5.3 Flash on a Mac

For most Mac users the practical route is server-side. The Coding Plan works with Claude Code, Cline and similar harnesses, and any multi-model client that supports GLM will reach the new model over the API without a download. Fello AI puts GLM alongside Claude, GPT, Gemini and Grok in one native Mac app, and the GLM desktop client for macOS page covers that setup.

Is GLM 5.3 Flash Worth Switching To

If you are paying frontier prices for agentic coding and automation work, this is the most interesting release of the month. This model lands within a point of Claude Opus 4.8 on Terminal Bench and beats it outright on AutomationBench, at a tenth of Z.ai’s own flagship pricing. That changes the arithmetic on any workload measured in millions of tokens.

Try it before the promotional rate ends on September 9.

If you want the best raw scores available, this is not that model, and the benchmark table above is honest about it. GPT-5.6 Terra leads three of the six tests outright.

And if you came for the open weights specifically, note carefully which ones arrived. GLM 5.3 Flash is downloadable today under MIT, which makes it the strongest openly licensed model Z.ai has published since GLM 5.2 and its 1M-token context. The 744B GLM 5.3 weights promised a fortnight ago are still not out, and the company has not said when they will be.

FAQ