GLM 5.3 is now available on Helmcode

The open ceiling, served on European infrastructure

On 25 August, Z.ai published the weights of GLM 5.3. On the Artificial Analysis Intelligence Index it scores 60, which is exactly what Kimi K3 scores. For the first time there are two models tied at the top of what you can download, and the important part is that one of them fits on a quarter of the hardware.

As of today, GLM 5.3 replaces GLM 5.2 on every Helmcode Enterprise account.

Same European infrastructure, same zero logs, same 150 euros a month per API key. None of that changes. The model does.

Two things get better:

• Seven points of intelligence. GLM 5.2 scored 53 on that index. GLM 5.3 scores 60.
• The full million tokens of context. We used to serve 500K of GLM 5.2's 1M window. On 5.3 we serve all of it.

If you already have a key with GLM 5.2, it answers on GLM 5.3 from today. There is nothing to migrate beyond the model id. And both will stay available for a month: that is how long Enterprise accounts have to move from GLM 5.2 to GLM 5.3.

GLM 5.3 in one minute

GLM 5.3 is built on the same base as GLM 5.2: a Mixture of Experts of around 744 billion parameters, of which roughly 40 billion activate per token, 78 layers, 256 experts with 8 active at a time. It keeps the two reasoning modes, High and Max.

What changed is the post-training. What moved is how long the model keeps working without losing the thread, and that is why the benchmarks it wins are the long ones and the ones it ties are the short ones.

Two numbers we read off the model's own repository:

• `max_position_embeddings` is 1048576. The million tokens of context is the model's own configuration.
• 755.6 GB of weights across 141 safetensors in native FP8, with a separate BF16 checkpoint at 1,506.7 GB. That is the number that decides how many GPUs anyone needs to run it.

One thing that did not change: GLM 5.3 is text only. No vision, no audio. The Hugging Face pipeline is text-generation and there is no vision configuration in the repository. If you need a model that reads images or scanned pages, that is Qwen3.6 or Gemma 4, and your plan already covers both at no extra cost. And very soon we will be adding GLM-5.3-Flash, which is multimodal.

The benchmarks

Everything in this section is Z.ai's own published table. Independent coverage of a model nine days old is still thin.

Against Kimi K3, the model it now ties on the index:

What we can read from that table:

On the short agentic benchmarks the two models are the same model. Except that GLM 5.3 is far cheaper to serve.

On the long benchmarks the gap is real and it favours GLM 5.3. Terminal Bench 3.0 is the harder, longer successor to 2.1, and 28.3 against 17.4 is a 60% lead. Humanity's Last Exam with tools, which is reasoning across many steps with something to call, goes 62.5 against 59.8.

K3 still wins Toolathlon. If your workload is heavy orchestration across many tools, measure both before you choose.

And for scale, against GLM 5.2: 28.3 against 4.6 on Terminal Bench 3.0. GLM 5.2 simply could not do that benchmark. That alone makes the swap worth it.

GLM 5.3 vs Kimi K3

When we launched GLM 5.2 we had to concede that Kimi K3 was better, and argue that GLM 5.2 was cheaper and very good. That argument is over. The two models now sit at a very similar score and a very similar level of performance.

Hardware. GLM 5.3 is 755 GB of weights. Kimi K3 is 1,561 GB. In practice that is roughly 16 GPUs of the H200 or B300 class against roughly 64 for K3.

Cost per task. Artificial Analysis measures 0.68 dollars per completed task for GLM 5.3 against 0.84 for K3.

Licences, and this one is closer than it was. GLM 5.2 was pure MIT against K3's own licence. GLM 5.3 is not pure MIT any more.

Vision. K3 reads images and video. GLM 5.3 does not. If that is your case, K3 genuinely wins. But we will soon be adding GLM-5.3-Flash to Enterprise so you have multimodal as a fallback.

The licence changed

What the LICENSE file contains is the MIT text word for word, plus one added condition: a licensee operating a Model as a Service business with more than 10 billion dollars of revenue over any twelve month period has to pass a Z.ai security review first.

It almost certainly does not affect you. If you are reading this as a company that wants to run or buy inference, the threshold is 10 billion dollars of revenue from a model serving business. That clause is aimed at four or five companies on earth.

It affects us in principle and not in practice. Helmcode is a Model as a Service business, so we are the kind of licensee the clause names. We are nowhere near 10 billion dollars.

Nobody can revoke it on you. The weights are published and the MIT grant underneath is intact. No one takes this model away from you, and no one starts logging your prompts retroactively.

What do I gain going through Helmcode instead of Z.ai's own API?

You use the best open model of the moment without your data going to Z.ai 's API. We serve it from our European infrastructure, we store no logs, and not a single token leaves our machines.

With AI Act penalties applying since 2 August, where and how your inference is served stopped being a technical detail some time ago.

The price

Unchanged. 150 euros a month per API key, as an add-on to any enterprise plan.

We said at the 5.2 launch that a model this size cannot go into the base plans without moving everybody's price, and that is still true: 755 GB of weights is not a rounding error on a GPU bill. So it stays opt in. Add as many keys with GLM 5.3 as you need, and only those keys cost extra.

Each key with GLM 5.3 includes 100 requests per minute, 5 concurrent processes, 3M tokens per minute, and context up to 1M per request.

How much those keys can consume in a month comes from your plan tier, as a token allowance shared across every key on the account. Unchanged as well:

When the allowance runs out, usage pauses until the next billing cycle. There are no overage charges: it is impossible to spend more than you contracted.

When does it pay off against paying per token?

We ran the numbers again, because the per token market moved since June. Assumption: a 3 to 1 input to output ratio, prices read on 3 September 2026, converted at the ECB reference rate of 2 September.

The simple answer: past roughly 100M tokens a month the rate pays for itself, and that is against the cheapest router we could find. Against Z.ai's own API it pays for itself at about 80M.

A single key burning the Starter allowance pays an effective €0.075/M, and a key that eats a whole Scale allowance drops below €0.01/M. No per token API reaches those numbers.

And the answer we would give you if you asked: if your volume is low and your data can travel to an American router, the router will come out cheaper. This product is for whoever consumes volume, needs the data in Europe, or needs to budget without shocks.

How to turn it on

It is available right now on the starter, growth and scale plans.

• If you already have a GLM 5.2 key, it answers on GLM 5.3 from today. Point the model id at glm-5.3. Your price, your limits and your allowance do not change.
• If you are a Helmcode customer without one, go into your dashboard and generate a new key. You will be able to assign it to GLM 5.3 and the cost is added to your next invoice.
• If you are not a customer yet, when you pick your plan the next step asks how many API keys with GLM 5.3 you need.

The endpoint is OpenAI compatible:

from openai import OpenAI

client = OpenAI(base_url="https://api.helmcode.com/v1", api_key="YOUR_KEY")
resp = client.chat.completions.create(
    model="glm-5.3",
    messages=[{"role": "user", "content": "..."}],
)

Full docs here: helmcode.com/docs

The best open model on the board is now the one on your plan, with the whole million tokens of context and the same bill as last month. Go and use it.