Separate orchestration, agent, model, and accelerator before sizing
Do not size “an AI VPS.” Size layers:
- Orchestration — n8n/schedulers (see n8n operations).
- Agent / UI — tools that can take privileged actions.
- Weights + KV — local model memory and disk.
- Accelerator / external API — GPU Servers/Spot or hosted inference bills separate from the Cloud Server.
| Architecture | Intent | Primary risk |
|---|---|---|
| Hosted-model agent on small Premium | API models; tiny always-on host | API spend + shut-down still bills |
| Local small CPU model on high-RAM Premium | Private weights; measured RAM/disk | OOM / disk / Fair Transfer pulls |
| Agent + external LLM | Privileged automation + hosted models | Credential blast radius |
| GPU Server / Spot path | VRAM needs beyond honest CPU experiments | Spot SLA exclusions; trial GPU unavailable |
| No-fit | Prohibited uses / expects pause=$0 on Premium | Do not deploy |
UpCloud hosting is not “local-only privacy” if prompts leave through external APIs.
Budget RAM and disk—not “7B VPS” labels
Inspect free RAM, swap, disk, and existing workloads before pulls. Worksheet fields: OS + runtime + download + resident weights + context/KV + concurrency reserve.
Editorial hypotheses only (not lab results, no tokens/s): consider 8 GB+ for one tiny model / one request if resources fit; consider 16 GB+ for a modest quantized 7B-class experiment if weights, KV, and OS fit; larger or context-heavy work → GPU Servers path or external inference. Do not paste uncaptured blog tutorials as AiXVPS fact.
Model pulls consume disk and Fair Transfer budget—monitor both. Premium shut down still bills; Native can stop compute while disks remain.
One bounded local-model reference on an inspected CPU host
Conditions: accepted Setup host; measured RAM/disk; Ubuntu/Docker as discovered.
- Prefer an AiXVPS-style project directory such as
/opt/aixvps/ollamawith an intentionally pinned image after you check current upstream docs. - Named volume; no blind
curl | sh; no privileged broad host mounts. - Default bind
127.0.0.1:11434or private network only; authenticated restricted remote access if needed; prove effective exposure. - Smoke: local API, list/show a
<SMALL_MODEL:PINNED_TAG>after license/disk/RAM check. - Cleanup named test artifacts only—never indiscriminate volume delete on a shared host.
Agents as privileged automation—not a toy image
Before connecting email, calendar, chat, or shell tools: inspect data paths, model providers, logs, and a kill/revoke plan. Use least privilege, human approval for dangerous actions, secrets outside git, and spend caps on external APIs separate from the VPS prepaid balance.
Calendar a delete for abandoned agents. Do not leave Native disks and IPs idle “until next week” without a balance alarm.
GPU Servers / Spot as exit—not the default five-article core
UpCloud documents GPU Servers and Spot paths. TOS language excludes Spot from SLA—say so. Docs free-trial quotas mark GPU unavailable—do not plan trial GPU. When VRAM or sustained inference exceeds honest CPU experiments, exit to GPU Servers, Spot (with SLA caveat), external inference, or another aixvps cluster with evidenced GPU guidance—without rankings.
No mining or AUP evasion tips.
Observe, stop, resize, split, or leave
| Symptom | Safer branch | Avoid |
|---|---|---|
| OOM / thrash | Shrink model/context; raise RAM; or leave to GPU/external | Ignore and add swap as a lifestyle |
| Disk full | Prune pulls; enlarge disk; Native attach carefully | Delete random volumes |
| API quota / $ burn | Caps; move offline; kill agent | Leave agent looping |
| Unexpected prepaid drain | Check Premium shut-down billing; leftover on-demand backups; Native disks | Assume pause=$0 |
| Fair Transfer pressure | Throttle pulls; change architecture | Invent overage $/GB |
Link billing before destroy. Link n8n for orchestration without publishing inference endpoints.