Server for Qwen
Configurations for Qwen and other open-weight models, sized to model size, context and concurrency.
What it is
The Qwen family spans from compact 3B models to MoE architectures with hundreds of billions of parameters, plus specialised variants: Coder, VL (vision), Math. We size infrastructure to the specific variant and scenario — not to “Qwen in general”.
When it fits
- Assistants with strong Russian and English support
- Coding assistant (Qwen-Coder)
- Vision tasks (Qwen-VL)
- Efficient MoE models on limited hardware
What to keep in mind
- MoE variants require special handling of memory and concurrency
- VL models need extra resources for image processing
- Quantising MoE is more complex than dense models
- Russian quality depends on the version and fine-tuning
Frequent questions
Qwen or Llama for Russian?+
Qwen generally performs better in Russian, especially recent versions. Llama remains a strong choice for English tasks. Final choice follows testing on your data.
What is MoE and should we choose it?+
MoE (Mixture of Experts) activates only a subset of parameters per request. That gives high quality at lower compute cost, but needs more memory. An appropriate runtime is essential to use it efficiently.
Can Qwen-VL run locally?+
Yes, but VRAM requirements are higher than for text models of the same size due to image processing. We size the configuration to the specific VL scenario.
Let’s discuss your task
Tell us what the AI contour needs to solve — we will match the configuration, infrastructure and next step.