VPS Specs and AI Model Limits

Every OpenClaw-powered ClawOps bot runs on a dedicated Hetzner CX23 VPS. This page explains what that means for model selection and what to do if your bot becomes unresponsive.

CX23 specs

Resource	Amount
CPU	2 vCPU (shared)
RAM	4 GB
Disk	40 GB SSD
Network	20 TB/month

This is a solid general-purpose VPS. It handles most AI models without issue, but very large models with unusual token specifications can cause problems (see below).

Which models work well

Models with a context window up to around 262K tokens run reliably on the CX23. This covers the vast majority of models available through OpenRouter, including:

• All four built-in model stacks (Wanderer, Hustler, Professional, Operator)
• Gemini 3.7 Flash and Gemini 3.5 Flash Lite (1M context, 65K output)
• GLM-5.3, Qwen 3.8 Max, DeepSeek V4 Flash (1M context, ~131K output)
• Claude Sonnet 5 / Opus 5 (1M context, 128K output)

The real test is context window minus maximum output tokens — how much room is left for your message and history after the model reserves its output budget. A model is fine as long as that’s comfortably large.

Models that don’t work

Some models — often newer 1M context window ones — report a max_output_tokens close to the full context window (for example 512,000 on a 524,288-token window, or ~900,000 on a 1M window). That leaves almost no room for input, so every message overflows immediately, even on a fresh session.

The catch: this varies by hosting provider. The same model can be safe on one OpenRouter provider and unusable on another. Examples where at least one provider is affected:

• MiniMax M3
• Kimi K3
• DeepSeek V4 Flash (safe on some providers, not others)
• Grok Build 0.1

Symptom: You’ll see this message in Telegram right after sending any message, even on a fresh /new session:

⚠️ Auto-compaction could not recover this turn.

This is not a VPS memory issue, it’s a model specification problem on the provider’s end. The fix is to switch to a different model.

Note

This affects the custom model picker only. The built-in model stacks route around the affected providers automatically and are checked per endpoint for CX23 compatibility before any changes are made.

What to do if your bot is slow or unresponsive

If your bot responds slowly or not at all after switching to a new model:

1. Try a different model, open Dashboard → Settings → Custom Model and switch to something with a smaller context window, or just switch back to a built-in stack
2. Restart the gateway, Dashboard → Settings → Restart Gateway
3. Run Repair & Update, Dashboard → Settings → Repair & Update. This checks for OpenClaw updates, fixes config issues, and restarts the gateway automatically
4. Start a fresh session, send /new to your bot in Telegram to clear any corrupted session state

Compaction

Your bot automatically compacts long conversations to keep them within the model’s context window. Compaction uses Gemini 2.5 Flash as a dedicated lightweight model, separate from your primary model, so it works reliably regardless of which stack or custom model you’re using.

If you see repeated compaction errors after switching models, use Repair & Update from the Dashboard to reset the configuration.

Upgrading your VPS

A VPS upgrade option (CX32 / CX42) is on the roadmap for users who want to run heavyweight 1M-context models reliably. If this is something you need, contact support@clawops.io.