The self-hosted coding stack

All four tools here have a genuine free tier, so a working configuration costs nothing while you evaluate it. The 03 COSTS table above breaks down each tool. Self-hosting shifts the meter off seats and tokens onto two things that climb with use: GPU capacity, and the engineering time to serve the weights. Hardware is the visible cost; the time to run it reliably is the one teams underbudget. Price the people, not just the GPUs.

Rarely. Qwen on Ollama covers the first move: a capable open model running locally on hardware you own, enough to prototype before anything is served. The numbered steps lay out the capability layers, not an adoption sequence. Start there, bring in vLLM when a prototype needs to become infrastructure other systems call, and add DeepSeek only when a task suits it better than Qwen.

Then you are already running the local end of this stack: Ollama gives you a model on your own hardware, no cloud and no account. Keep it for prototyping. What the stack adds is the path off the laptop: vLLM serves the same open weights at production throughput for other systems to call, Qwen sets the default model, and DeepSeek is the second option. Ollama proves the idea; vLLM runs it.

vLLM and Ollama both run open weights on infrastructure you control, the one overlap here. The dividing rule is audience. Ollama serves one developer at a time: a clean command line on ordinary hardware, made for prototyping. vLLM serves everything else: production throughput for the systems that call the model. The vLLM and Ollama comparison, linked above, covers where the line falls.

Often, and this tier says so. Self-hosting earns its keep only when code cannot leave your own infrastructure and you have the GPUs and engineering time to serve models well. Two signals say you have overshot: no policy or jurisdiction rule forbids a hosted open-weight API, or nobody owns the serving infrastructure. A team reaching for control it does not need wants the competitive tier, the open-weight coding agent stack.

Self-hosting inverts the usual answer: run these weights on infrastructure you own and confidential code stays there. Two conditions hold that guarantee. Keep Ollama's optional web-connected features off, since one outbound call breaks the promise; and vLLM's security is now yours to own, not a vendor's. The line not to cross is the hosted APIs: Qwen's runs on Alibaba Cloud and DeepSeek's under Chinese jurisdiction, so sensitive material belongs only on the weights you host yourself.