# Self-Hosted LLM on a $5 VPS in 2026: What Actually Works > Source: > Published: 2026-08-12 05:05:12+00:00 "Run your own ChatGPT for five bucks a month" sounds like cheap clickbait. And mostly it is. But the gap between clickbait and reality has narrowed a lot in 2026. Quantized 3B and 7B models have become genuinely useful. VPS providers now pack 8 GB of RAM into plans under \$6. If you pick carefully, you can run a respectable local model on a budget server. This article is about what actually works, what falls over, and which provider from our [hosting comparison](https://hostingsift.com/hosting) gives you the most compute per dollar for AI workloads. Three reasons keep coming up in our research: **Privacy.** 44% of organizations cite data privacy as the top barrier to adopting hosted LLM APIs. Self-hosting means your prompts, documents, and customer data never leave your box. For law firms, clinics, and anyone building internal knowledge bases, this is non-negotiable. **Cost at volume.** Hosted API costs add up fast once you run batch jobs, background summarization, or embedding pipelines. The rough break-even with a \$5 VPS sits around two million tokens per day for small models. Below that, APIs are cheaper. Above that, self-hosting starts to win. **Learning and control.** You get to pick the model, tune the system prompt at the weight level (via LoRA adapters), and run without rate limits. Worth it for developers. Before we compare VPS plans, a quick hardware reality check. On CPU-only servers (which every budget VPS is), the bottleneck is RAM bandwidth, not CPU speed. Model weights must fit in RAM, and every token generation pass reads the entire weight file. So the two numbers that matter are: Here is roughly what each tier can do in 4-bit quantization (GGUF format, llama.cpp inference): | RAM | Largest usable model | Realistic tokens/sec (CPU) | Good for | |---|---|---|---| | 2 GB | Qwen 2.5-0.5B, Llama 3.2-1B | 20-40 tok/s | Embeddings, simple classification | | 4 GB | Llama 3.2-3B, Phi-3.5 Mini | 8-15 tok/s | Personal chatbots, summarization | | 8 GB | Qwen 2.5-7B, Mistral-7B, Llama 3.1-8B | 4-8 tok/s | Production-adjacent use, RAG | | 16 GB | Mixtral 8x7B (barely), Qwen 2.5-14B | 2-4 tok/s | More complex reasoning | Below 8 GB you are limited to small models. Above 8 GB you unlock the "actually useful" tier. This matters enormously when picking a VPS. Pulled directly from our database, filtered to plans where the RAM is at least 8 GB (the practical floor for running a 7B model in 4-bit quantization): | Provider | Plan | Price / mo | RAM | vCPU | Storage | Bandwidth | |---|---|---|---|---|---|---| | Compare that to the "famous" \$5 cloud plans: | Provider | Plan | Price / mo | RAM | vCPU | |---|---|---|---|---| | That 8x RAM difference at the same price point is the whole reason European providers win this category. DO, Vultr and Linode are excellent for web apps and small services. They are not designed for the RAM-heavy workload an LLM needs. Contabo Cloud VPS 10 at \$3.60 gives you 8 GB RAM, 4 vCPU and 75 GB NVMe. Nothing else in the market matches this at that price. The catch is real: Contabo is known for inconsistent CPU steal on noisy neighbors, and network latency can be variable. For LLM inference (which is RAM-bound, not network-bound) that tradeoff is acceptable. Go up one tier to the Cloud VPS 20 at \$5.60 for 12 GB RAM and 6 vCPU. This is the sweet spot for running Qwen 2.5-14B in 4-bit and still having headroom for a web server and a small vector database on the same machine. The Hetzner CX33 at \$5.49 with 8 GB RAM and 4 vCPU is the most reliable choice in this range. AMD EPYC cores, fast NVMe, 20 TB of bandwidth, datacenters in Germany, Finland and Virginia. Support is competent, the control panel is minimal in a good way. The ARM variant (CAX21) at \$6.49 is especially interesting for LLM work. Ampere Altra cores benchmark about 15 to 25 percent faster than the x86 CX33 on llama.cpp, because of higher sustained memory bandwidth. Same price range, noticeably better tokens per second. Less known outside the German-speaking market. The VPS 1000 ARM G11 at \$6.26 gives you 8 GB RAM, 6 vCPU, and 256 GB of NVMe, which is the largest disk in this comparison. Useful if you want to store multiple models and switch between them without re-downloading 4 to 8 GB each time. Here is the honest menu for an 8 GB VPS with Ollama or llama.cpp: What you will *not* run well: anything 14B+, models with vision capabilities (qwen-vl, llava), or anything demanding low-latency streaming to many users. Rough sketch so you know what you are signing up for. This assumes Ubuntu 22.04 or 24.04 on a 8 GB VPS. ``` # 1. Install Ollama (handles download, quantization, serving) curl -fsSL https://ollama.com/install.sh | sh # 2. Pull a model ollama pull qwen2.5:7b-instruct-q4_K_M # 3. Test ollama run qwen2.5:7b-instruct-q4_K_M "Summarize the French Revolution in two paragraphs." # 4. Expose the API (careful with firewall rules) # Ollama serves on localhost:11434 by default. Reverse-proxy with Caddy or Nginx. ``` Add Open WebUI if you want a ChatGPT-like browser interface. Add `llama-swap` if you want to load different models on demand without restarting. A \$5 VPS gives you about 5 to 8 tokens per second on a 7B model. That is fine for one user typing in a chat. It is not fine for: For those workloads, you want a GPU instance. That immediately pushes you above \$100/month with Vast.ai, RunPod, or a dedicated server with a consumer GPU. At that point, OpenAI API or Anthropic API is often cheaper up to surprisingly high volumes. | Use case | Best pick | Why | |---|---|---| | Lowest price with 8 GB RAM | Contabo Cloud VPS 10 | \$3.60 for 8 GB is unmatched | | Most reliable inference | Hetzner CX33 | Predictable performance, no noisy neighbors | | Best tokens per second per dollar | Hetzner CAX21 (ARM) | Higher sustained memory bandwidth | | Largest model storage | Netcup VPS 1000 ARM G11 | 256 GB NVMe stores a dozen models easily | | Running LLM + web app on one box | Contabo Cloud VPS 20 | 12 GB RAM leaves room for Postgres and nginx | You can also compare them side by side on [Contabo vs Hetzner](https://hostingsift.com/compare/contabo-vs-hetzner) or review the full Two developments worth tracking in 2026: **Smaller models keep getting smarter.** Phi-3.5 Mini already outperforms last year's 13B models on some benchmarks. Expect 2B and 3B models to be genuinely production-useful by Q4 2026. That makes the 2 GB RAM tier (under \$2) a legitimate option. **ARM inference is closing the gap.** Apple showed what M-series can do. Ampere Altra and AWS Graviton are bringing similar memory bandwidth advantages to \$5 VPS plans. If you have not tried ARM yet, this is the year. If you want to learn, build a personal assistant, run background summarization, or stand up a private chatbot for a small team, a \$5 to \$7 VPS with 8 GB of RAM is genuinely enough in 2026. Pick [Contabo](https://hostingsift.com/hosting/contabo) for raw price, [Hetzner](https://hostingsift.com/hosting/hetzner) for reliability, [Netcup](https://hostingsift.com/hosting/netcup) for storage. Avoid the US hyperscaler budget tiers, they are optimized for a different use case. If you want to serve real users at real volume, stop pretending a shared VPS is the answer. Use an API, or skip straight to a GPU host. *Originally published at hostingsift.com, where we track hosting prices nightly across 50 providers.*