{"slug":"cloudscale-ch-evaluates-self-hosted-coding-llms-on-its-own-gpu-instances-with-vllm-and-litellm","url":"https://findausecase.com/use-cases/cloudscale-ch-evaluates-self-hosted-coding-llms-on-its-own-gpu-instances-with-vllm-and-litellm","title":"cloudscale.ch evaluates self-hosted coding LLMs on its own GPU instances with vLLM and LiteLLM","description":"Swiss cloud provider cloudscale.ch spent several weeks evaluating self-hosted coding models on its own GPU instances (NVIDIA RTX PRO 6000 Max-Q, 96GB), served with vLLM behind a LiteLLM gateway and used with agents such as opencode, compared against Claude Opus 4.6/4.7 via Claude Code and Swiss-hosted models from Infomaniak. The company concludes self-hosting is viable but not plug-and-play, and has not reached a final conclusion; the test was a single engineer, with a production-like test still pending.","company":"cloudscale.ch","industry":"Technology & Software","country":"Switzerland","aiCapabilities":["Generative AI","Agentic AI"],"technology":["vLLM","LiteLLM","Claude Code"],"deployment":"Unknown","problemStatement":"Engineers and businesses increasingly expect to develop software with the help of coding agents. cloudscale.ch cares about data privacy and digital sovereignty and has its own GPU instances, so it wanted to compare self-hosting coding models with hosted models and public providers such as Anthropic's Claude Code.","solutionApproach":"cloudscale.ch ran coding models on one VM with a single NVIDIA RTX PRO 6000 Max-Q GPU (96GB VRAM), served by vLLM behind a LiteLLM API gateway with Caddy for TLS, and used them with agents such as opencode. It compared self-hosted models (Qwen2.5-Coder-32B, deepseek-coder-33b-awq, Gemma 4, Qwen3.6-35B) with Claude Opus 4.6/4.7 via Claude Code and with Swiss-hosted Infomaniak models (Qwen3-VL-235B, GPT-OSS-120B, Kimi-K2.5).","businessValue":"The source reports no quantified outcomes. It concludes that self-hosting coding LLMs is viable but not plug-and-play: it gives full control, data privacy and predictable cost, but requires time to tune, debug and maintain. The test was minimal (a single engineer), a production-like test is still pending, and cloudscale.ch has not yet reached a final conclusion.","evidence":{"band":"high"},"sourceUrl":"https://cloudscale.ch/de/engineering-blog/2026/05/04/self-hosting-coding-llms","dates":{"publishedAt":"2026-10-21T05:47:30.587Z","publishedAtSource":"pipeline","updatedAt":"2026-10-21T05:47:30.587Z"},"license":"Open for reading and citing with a link to https://findausecase.com/use-cases/cloudscale-ch-evaluates-self-hosted-coding-llms-on-its-own-gpu-instances-with-vllm-and-litellm. Bulk republication requires permission."}