Open weights give you options. They do not run themselves.
Models such as Qwen3.8 make self-hosting a practical subject for evaluation. The software licence is only one part of the operating decision.
A model you can evaluate on your own terms
The official Qwen3.8-27B model card provides a concrete example of a downloadable model with an Apache 2.0 licence. Its availability lets a team investigate its own deployment, subject to the model's documented requirements and terms. It does not establish that the model is the right choice for every application or that it will fit every GPU. Qwen: official Qwen3.8-27B model card ↗
This is the useful promise of open weights: an organisation can evaluate the model itself and retain more control over how it is deployed. That control has value when the workload, data handling or customisation needs justify the engineering effort.
Start with quality, then size the machine
Choose an evaluation set that reflects the work the model will perform. A document-processing service needs different checks from a coding assistant or a visual-inspection application. Include failure cases, not only examples that are easy to demonstrate.
Once the quality is acceptable, measure the resources needed at the desired service level. Model weights are part of the memory requirement; runtime allocations, context and concurrent requests add to it. Quantisation can change capacity and performance, but its effect on output quality needs evaluation too.
Avoid promising a user count from model size alone. Two services using the same model can have very different input lengths, output lengths and response-time requirements.
Ownership includes the operating work
A self-hosted deployment needs updates, monitoring, backups where applicable and a recovery procedure. Someone must maintain the environment and decide when a new model version is ready to replace the current one. A downloadable checkpoint does not provide those services by itself.
There is also a boundary to any privacy claim. Running inference locally can reduce external processing, but application connectors, telemetry and other services may still transmit information. Review the complete data path before describing a deployment as local or private.
Keep the comparison fair
A managed API may be a good choice for irregular demand or a small team that wants to avoid infrastructure work. An owned system may suit sustained demand, a specialised workflow or a requirement for direct control. Both assessments depend on actual usage and operating responsibilities.
Compare the cost of an acceptable completed task, including the people and supporting services needed to deliver it. Keep an exit plan: preserve evaluation data, model settings and deployment definitions so that changing the model or hardware remains possible.
Open weights expand the available choices. Their strongest commercial advantage is the ability to test and control a deployment, rather than an assumption that downloadable software makes inference free.
Sources & further reading
Primary sources for the reported developments and technical context. Analysis and conclusions are our own; linked specifications and documentation can change.
Sources checked 29 September 2026.
