Qwen 3.8 27B will be available to everyone for one week

At QDivZero, we are entering the final stretch of our private beta. Over the past few months, we have expanded the platform, added new models and, above all, learned a great deal about how inference infrastructure behaves when it moves beyond our benchmarks and starts handling real-world workloads.

Every now and then, we do something that has proven particularly useful and that we know many of you look forward to. We temporarily open one of our models to everyone. This time, the choice was fairly obvious. Qwen 3.8 27B is one of the models you have been asking us for the most and, since your wish is our command, it will be available to everyone on QDivZero for the next seven days.

Why Qwen 3.8 27B

At 27 billion parameters, Qwen 3.8 27B sits at a particularly interesting scale. It has the capacity to handle demanding tasks across programming, reasoning, analysis and text generation, while presenting an inference profile that allows us to study how our infrastructure behaves under very different workloads.

Real-world model performance is about much more than tokens per second. Once multiple users, different context lengths and simultaneous generations enter the picture, batching, scheduling, memory utilization, KV cache management and the relationship between throughput and latency all start to matter.

A configuration that performs extremely well in a benchmark can behave very differently once real traffic arrives. That behavior is precisely what we are interested in observing.

A week of real traffic

We can simulate concurrency and build internal benchmarks, and we do so constantly. There is, however, an important difference between a workload designed to measure a system and one generated by people who simply want to use it.

This week, some of you will use Qwen for coding, others will work with long documents, build agents or simply talk to the model and experiment with it. For us, that diversity provides particularly valuable information.

We want to see how the optimizations we have developed over the past few months behave once conditions stop resembling our own test scenarios. We are approaching the end of the private beta, which makes this a good time to broaden the traffic again and understand the limits of the system more precisely.

Privacy as a principle

One thing does not change with this public opening. User privacy is one of the principles on which we are building QDivZero.

We will not log the content of your interactions with the model. What you write, the code you test and the documents you work with are not what we want to learn from this week.

We are interested in understanding how the infrastructure behaves, not in knowing what you are saying to the model.

We want you to be able to use it freely, experiment and bring real use cases knowing that the content of those interactions will not be logged. Making access public for a week does not change the way we approach privacy.

There is much more behind QDivZero

Qwen 3.8 27B is only a small part of what we have available. QDivZero has a much broader model catalogue, including models available through Hugging Face as well as models from other sources. During the private beta, however, that full catalogue is not open to the general public.

These temporary openings are our way of sharing a small part of it every now and then. For a few days, we make one of those models public so that anyone can try it, experiment with it and tinker without needing to be part of the beta.

These openings are useful to us because they introduce new usage patterns and help us learn more about how the platform behaves. But not everything needs a technical explanation. We know you enjoy trying different models, and we have plenty that we can occasionally share with you.

Choosing the next one does not have to be entirely up to us either. If there is a model you would like to use on QDivZero, we want to hear about it. Think of these openings as small gifts for those of you following along while we continue building the platform.

You have seven days

Qwen 3.8 27B will be available to everyone on QDivZero for the next seven days.

The Qwen 3.8 27B demo beta has ended. More experiences are coming soon.

You can use it for coding, reasoning, agents, documents, long-context experiments or that idea you have been waiting to test with the right model. We will use the opportunity to learn how the system behaves as we approach the end of our private beta.

There will be more models and more opportunities to experiment with parts of the catalogue that normally remain inside QDivZero.

If there is one you would like us to open next, tell us which one.

Your wish is our command.