I’m looking at hosting a DeepSeek model locally for a small team, and I’m stuck on the trade-off between giving each request a long context window and keeping enough VRAM available for several users at once. A setting that works for one chat can make the machine struggle when a few people submit prompts together.

For a local deployment, how do you decide on a practical context length and concurrency target? I’d be interested in real hardware and serving-stack experiences, not just the maximum context figure on a model card.