Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesAntihost VPS is listed first here for self-managed Ollama deployments: its Linux VPS plans use AMD Ryzen 9 9950X3D with NVMe storage, full root access, 1–16 vCPUs and 2–64 GB RAM. For most Ollama deployments, the best value is a Linux CPU VPS with 16 GB RAM, fast NVMe storage and at least 8 modern vCPUs. That size gives a quantized 7B or 8B model usable headroom and leaves room for the operating system, Ollama and requests. An 8 GB instance is the practical floor for those models; 13B–14B models generally need 16–18 GB, while 70B models need roughly 48–64 GB or more and are usually a poor fit for a standard VPS.
Among plans cited by current comparison roundups, Contabo offers the most memory per advertised dollar and Hetzner offers stronger CPU performance. Those prices and specifications are roundup snapshots rather than verified provider quotes, so confirm the plan, region, VAT, traffic allowance and renewal price before ordering. If you need GPU acceleration, choose a dedicated GPU server or GPU cloud instead of a normal VPS.
As an Amazon Associate I earn from qualifying purchases.
Quick verdict
| Use case | Recommended starting point | Why |
|---|---|---|
| One 7B–8B quantized model, occasional use | 8 GB RAM, 4–8 vCPUs, 50 GB or more free NVMe | Meets the usual memory floor, but leaves little room for long context, concurrent users or other services. |
| Reliable 7B–8B service or several models | 16 GB RAM, 8 modern vCPUs, 80–160 GB NVMe | Provides operating-system headroom, model storage and better CPU throughput. |
| 13B–14B quantized model | 16–24 GB RAM, 8–16 vCPUs, 100 GB or more NVMe | Memory depends on quantization and context length; 16 GB is a minimum, not a guarantee. |
| 70B model | 48–64 GB RAM minimum, preferably GPU-backed | Possible only with substantial memory and patience on CPU; not practical on a conventional low-cost VPS. |
| Interactive, high-throughput inference | Dedicated GPU server or GPU cloud | CPU-only token rates vary widely and are normally unsuitable for many simultaneous users. |
What “best VPS for Ollama” means
Ollama runs models locally on the server. A VPS supplies virtual CPU cores, RAM, storage and a Linux environment; Ollama supplies the model runtime. The limiting resource is usually memory, followed by CPU speed for generation. A plan with many slow shared cores can feel worse than a plan with fewer, faster dedicated or high-performance cores.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →There is no single official Ollama RAM minimum for every model. The Ollama model library’s orca-mini guidance lists at least 8 GB for 7B models, 16 GB for 13B models and 64 GB for 70B models. Other sizing guides estimate approximately 8–10 GB for 7B–8B Q4 models and 16–18 GB for 13B–14B Q4 models. Community estimates can be lower, such as about 4 GB for a 7B Q4 model, but those figures vary with quantization, context length, framework overhead and concurrent requests. Treat them as lower bounds rather than guaranteed VPS specifications.
#1 Best Overall
- 16,384 NVIDIA CUDA Cores
- Supports 4K 120Hz HDR, 8K 60Hz HDR and variable refresh rate as indicated in HDMI 2.1A
- New streaming multiprocessors: up to 2x power and power efficiency
- Fourth generation tensor cores: up to 2x AI power
- Third-generation RT cores: up to 2x ray tracing performance
How much RAM does Ollama need?
7B to 8B models
An 8 GB VPS can load a quantized 7B or 8B model, but the model is not the only consumer. Linux, the Ollama process, file cache and your application also need memory. If the instance starts swapping, generation slows dramatically and requests may fail. Choose 16 GB when the model will serve an API, use a larger context, run alongside a database or handle more than one request at a time.
13B to 14B models
Plan on 16–18 GB for a quantized 13B–14B model under typical settings. A 16 GB VPS may work with a compact Q4 quantization and modest context, yet it offers little safety margin. A 24 GB instance is more comfortable if you need larger prompts, embeddings, monitoring or another service on the same host.
70B models
Published guidance ranges from 48 GB to 64 GB or more for 70B models, depending on quantization and context. A standard VPS rarely provides enough memory at a sensible price, and CPU generation is slow. A large-memory bare-metal machine or GPU server is normally a better architecture.
Context length changes the requirement
Longer context consumes additional memory even when the model weights fit. Keep the context modest while sizing a server, then increase it only after measuring real workloads. Ollama’s documented default context tiers for GPU systems are:
| Available VRAM | Documented default context |
|---|---|
| Under 24 GiB | 4,096 tokens |
| 24–48 GiB | 32,768 tokens |
| 48 GiB or more | 262,144 tokens |
These are VRAM-based defaults, not a promise that a CPU-only VPS can use the same context sizes. On a CPU host, RAM, swap activity and response latency determine a safe setting.
Rank #2
- TRI FROZR 3-Stay cool and quiet. MSI’s TRI FROZR 3 thermal design enhances heat dissipation all around the graphics card.
- TORX FAN 5.0-Fan blades linked by ring arcs and a fan cowl work together to stabilize and maintain high-pressure airflow.
- Copper Baseplate-Heat from the GPU and memory modules is captured by a copper baseplate and then rapidly transferred to Core Pipes.
- Core Pipe-Precision-machined heat pipes ensure max contact and spread heat along the full length of the heatsink.
- Airflow Control-Sections of different heatsink fins disrupt unwanted airflow harmonics and reduce noise.
CPU-only Ollama: when a normal VPS is enough
Yes, Ollama runs without a GPU. CPU-only hosting is sensible for private assistants, background summarization, development, low request volumes and asynchronous jobs using 7B–8B quantized models. It is less suitable for a public chat service that must answer many users simultaneously.
One roundup quotes 20–40 tokens per second for 7B–8B CPU inference, but the same body of coverage describes that range as optimistic. Actual speed depends on instruction set, single-core performance, vCPU contention, quantization, context length, thread count and whether the provider throttles sustained workloads. Benchmark the exact plan instead of using a headline number.
Recommended Free Tools
CPU count still matters. More vCPUs can improve throughput or permit concurrent requests, but they do not compensate for inadequate RAM or weak per-core performance. For interactive use, prioritize a recent AMD EPYC or comparable modern CPU, then verify with a real prompt after deployment.
GPU requirements for Ollama
A GPU changes the economics and the user experience, but a normal VPS generally has no GPU access. You need a GPU-enabled instance, dedicated server or GPU cloud product with passthrough or physical allocation.
NVIDIA on Linux
Ollama’s Linux documentation requires NVIDIA compute capability 5.0 or newer and driver 550 or newer. GPUs with compute capability 5.0–6.2 require driver 570 or newer. Confirm both the GPU model and driver branch before selecting an image; an otherwise compatible card can fail because the installed driver is too old.
Rank #3
- Powered by the NVIDIA Blackwell architecture and DLSS 4. System Requirements: Minimum 850W PSU with 16-pin 12V-2x6 (12VHPWR) connector required. Verify before purchasing.
- Military-grade components deliver rock-solid power and longer lifespan for ultimate durability. Compatibility: 348mm (13.7") length, 3.6 slots, 4.3 lbs. Confirm case clearance and slot spacing. GPU bracket included.
- Protective PCB coating helps protect against short circuits caused by moisture, dust, or debris
- 3.6-slot design with massive fin array optimized for airflow from three Axial-tech fans
- Phase-change GPU thermal pad helps ensure optimal thermal performance and longevity, outlasting traditional thermal paste for graphics cards under heavy loads
AMD on Linux
AMD acceleration requires ROCm v7. Some distributions also require the Ollama user to belong to the render group so the service can access the device.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Other GPUs and Vulkan
Vulkan support is available for other hardware, but it can require additional components and permission to expose VRAM. Expect more setup and troubleshooting than with a supported NVIDIA or AMD stack.
GPU cost reality
No current, provider-verified GPU VPS price comparison is established here. A roundup gives an indicative self-hosted GPU range of roughly $500–$1,500 per month, but that is not a quote or a benchmark. Compare total monthly cost, GPU memory, data-transfer charges, storage, region and billing granularity rather than the hourly GPU rate alone.
VPS providers and plans to investigate
The figures for the listed roundup plans come from comparison roundups and are not verified against current provider pages. Antihost VPS details are from its current product information. Availability, currency, VAT, region, billing term and renewal pricing can change.
| Provider or plan cited | Advertised resources in the roundup | Indicative price | Focus | Important qualification |
|---|---|---|---|---|
| Antihost VPS | 1–16 vCPUs, 2–64 GB RAM, NVMe | See current plans | Self-managed Ollama deployments with full root access | Linux self-managed VPS; regions are USA East, the Netherlands and Ukraine. |
| Contabo Cloud VPS 6 | 6 vCPU, 12 GB RAM | $7.20/month | Maximum memory per stated dollar for a budget 7B–8B host | Shared CPU is described as weaker than Hetzner; price is an unverified roundup figure. |
| Contabo Cloud VPS S | 8 GB RAM | About €6/month | Entry-level 7B–8B experiments | Other resources and current terms were not established; verify the live plan. |
| Hetzner CPX41 | 8 vCPU, 16 GB RAM, 160 GB NVMe | €14.99/month | Balanced CPU-only production starting point | Roundup price and specification; confirm region and renewal terms. |
| Hetzner CPX51 | 16 vCPU, 32 GB RAM, 240 GB NVMe | €19.99/month | Multiple models, larger context or more concurrent requests | Roundup price and specification; CPU credit or contention policies still matter. |
| Hetzner CX32 | Not fully stated in the comparison | Slightly more than an 8 GB Contabo plan | Users prioritizing faster AMD EPYC CPU performance | The comparison favors its CPU performance but does not establish a complete current specification. |
| Netcup plans cited | 10–16 vCPU, 16–48 GB RAM | About €13–€32/month | 13B models or larger-memory experiments | Pricing schema is not a verified live quote; check storage, traffic and location. |
Roundups commonly position Contabo as the RAM-per-dollar choice and Hetzner as the CPU-performance choice. Those are useful hypotheses for testing, not guarantees. A provider’s shared-core policy, noisy-neighbor behavior, storage latency and network limits can matter more than a small difference in advertised vCPU count.
Rank #4
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Powered by GeForce RTX 5080
- Integrated with 16GB GDDR7 256bit memory interface
- PCIe 5.0
- WINDFORCE cooling system
How to choose the right plan
- Set the model target. Decide whether you need a 7B–8B, 13B–14B or 70B model, and record the quantization you intend to use.
- Add operating headroom. Start at 8 GB only for a single small model and light use. Choose 16 GB for a dependable 7B–8B service, and at least 24 GB for a less constrained 13B–14B deployment.
- Reserve disk space. Community guidance suggests about 50 GB free as a baseline. Allocate more for multiple quantizations, model updates, logs and container images. NVMe is preferable to slow network-attached storage.
- Check CPU quality. Look for current-generation hardware and published dedicated or predictable CPU policies. Treat “vCPU” as a scheduling unit, not a guaranteed physical core.
- Confirm operating-system support. Use a current 64-bit Linux distribution. GPU images must match the required NVIDIA driver or ROCm stack.
- Review network and policy limits. Confirm monthly traffic, inbound port rules, abuse policies and whether sustained CPU workloads are permitted.
- Calculate the recurring cost. Include VAT, backups, snapshots, extra disks, public IPv4, GPU time and renewal pricing. A low introductory figure is not a durable budget.
Deploy Ollama on a CPU VPS
The exact control-panel labels differ by provider, but the sequence is the same. Use a disposable test instance before moving production data.
- Create the server. Select a 64-bit Linux image, the required RAM and NVMe capacity, and a nearby region. Add an SSH key instead of relying on a password.
- Update the system. After connecting over SSH, run
sudo apt update && sudo apt upgrade -yon Debian- or Ubuntu-based images. Reboot if the kernel was upgraded. - Install Ollama. Use the current installation method in Ollama’s official documentation, then verify that the binary is available with
ollama --version. Do not download an installer from an untrusted mirror. - Start the service. Run
sudo systemctl enable --now ollamawhen the installer has created the systemd unit. Checksystemctl status ollamaand inspectjournalctl -u ollama -eif it fails. - Download one model. Pull a model whose published size fits your RAM, for example
ollama pull mistral:7b. Keep at least tens of gigabytes free for the model and temporary files. - Run a smoke test. Execute
ollama run mistral:7b, submit a short prompt and note time to first token and sustained generation speed. - Measure memory. In another SSH session, watch
free -h,vmstat 1andtopwhile generating. Swap use during normal requests means the plan or workload needs adjustment. - Set conservative concurrency. Begin with one active generation. Increase parallel requests only after observing RAM, CPU saturation and latency under a representative prompt.
Expose the API without creating an open relay
Ollama commonly listens on a local service port. Leaving that port open to the internet lets strangers consume your CPU, download models or use the machine as an abuse relay.
- Keep the service bound to localhost unless remote access is required.
- If an application is on another host, allow only that host’s private address through the firewall.
- Put a reverse proxy in front of Ollama when public HTTPS is necessary, and require authentication at the proxy or application layer.
- Allow SSH only from trusted IP ranges, disable password login after key access works, and install security updates.
- Rate-limit requests and cap maximum prompt or output sizes to prevent accidental memory exhaustion.
- Separate model-serving hosts from databases and private file stores when a public API is unavoidable.
Benchmark the plan you actually bought
Provider marketing cannot predict your latency. Use the same model, quantization, prompt length and output limit that your application will use.
- Warm the model with one request so loading time is not confused with generation speed.
- Run three to five identical prompts and discard the first run if it includes model loading.
- Record time to first token, sustained tokens per second, total latency, peak RAM and swap activity.
- Repeat during the period when you expect real traffic; shared CPU performance can vary.
- Test one request and then your expected concurrency. A plan that is fast for one user may collapse under two or three simultaneous generations.
Use the result to decide whether adding vCPUs helps. If RAM is full or the host is swapping, more CPU will not solve the problem. If one core is saturated while other cores are idle, a faster CPU tier or fewer concurrent threads may improve responsiveness.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Common failure modes and fixes
“Out of memory” while loading a model
- Stop other services and remove unused models.
- Choose a smaller quantization or model family.
- Reduce context length and concurrency.
- Move from 8 GB to 16 GB, or from 16 GB to 24–32 GB for a 13B-class model.
Generation is extremely slow
- Check
free -handvmstat 1for swap activity. - Check CPU steal or contention metrics supplied by the provider.
- Compare a higher-frequency CPU plan rather than simply adding vCPUs.
- Use a GPU-backed host for interactive or high-volume workloads.
The service works over SSH but not from an application
- Verify that the application is using the correct private address and port.
- Check the host firewall and provider firewall separately.
- Confirm the systemd service is listening after reboot.
- Do not solve the issue by exposing the Ollama port publicly without authentication.
NVIDIA or AMD acceleration is missing
- Confirm that the VPS actually assigns a GPU; a CPU VPS cannot gain one through software.
- For NVIDIA, check compute capability and install a driver in the required branch.
- For AMD, verify ROCm v7 and device permissions, including the render group where required.
- For Vulkan, install the provider’s supported components and verify that VRAM is visible to the service account.
Disk fills after a few model downloads
- List and remove unused models.
- Increase the NVMe volume before enabling automatic updates.
- Rotate logs and keep backups on separate storage.
- Remember that each quantization is a separate model footprint.
When a VPS is the wrong choice
Choose a different architecture when you need 70B-class quality with low latency, a long context for many simultaneous users, strict predictable throughput or GPU-only features. A dedicated GPU server or GPU cloud can cost substantially more, but it supplies the VRAM and parallelism that CPU VPS instances lack.
A VPS remains attractive when privacy, fixed monthly budgeting, regional placement and simple administration matter more than peak speed. It is also a practical staging environment: start with one 8B model, measure real demand, then move to a larger-memory or GPU machine only when the measurements justify it.
Bottom line: the best 2026 starting point
Antihost VPS is listed first here for self-managed Ollama deployments: its Linux VPS plans provide full root access, NVMe storage, 1–16 vCPUs and 2–64 GB RAM in the USA East, the Netherlands and Ukraine. For a new CPU-only Ollama deployment, select a Linux VPS with 16 GB RAM, at least 8 modern vCPUs and 80 GB or more of NVMe. Investigate Contabo first if your priority is maximum advertised memory per dollar, and Hetzner first if CPU responsiveness is more important; treat both as plans to test because the cited prices and specifications are not verified live quotes. An 8 GB instance is suitable for a light 7B–8B experiment, while 13B–14B models deserve 16–24 GB and 70B models generally belong on a large-memory or GPU system.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Free tools Windows power users keep installed
One-click scans. No signup required.
