Skip to content

STORAGE.md

Latest commit

 

History

History
135 lines (99 loc) · 5.34 KB

STORAGE.md

File metadata and controls

135 lines (99 loc) · 5.34 KB

Storage and download sizing

code-stick writes three big things onto the USB: engine binaries (Ollama + opencode, per OS/arch), model weights (data/), and a small config tree. This guide is about planning free space before you run code-stick install.

For flag syntax, see COMMANDS.md. For model sizes, see MODELS.md.

What the default install downloads

With no --targets flag, the installer stages all six platform trees so the same stick boots on Windows (x64 + ARM64), macOS (Apple Silicon + Intel), and Linux (x64 + ARM64):

Component Count on stick Approx. size
Ollama archives (extracted) 6 ~4–5 GB
opencode archives (extracted) 6 ~0.5–1 GB
Binaries total ~5–7 GB

The Windows and Linux x64 bundles are the heavy ones — each ships large CUDA/ROCm runtime libraries inside the Ollama zip. macOS and ARM64 trees are smaller, but you still pay for every target you stage.

On top of binaries you need:

  • Model blob(s) — see the Size column in MODELS.md (~2.2 GB for Phi-3 Mini up to ~20 GB for 32B tiers).
  • ~1 GB headroom — manifest, launchers, opencode config, temp files during install, and filesystem overhead.

USB size cheat sheet

Formatted capacity is always less than the label (an “8 GB” stick is usually ~7.2–7.5 GB usable). Plan on the usable number, not the marketing size.

USB label Default install (all 6 targets) Recommended approach
8 GB Does not fit for a usable stick --targets host + smallest model (phi3-mini, ~2.2 GB). See below.
16 GB Tight; only small models with full portability Full portable + Phi-3 or DeepSeek 6.7B; or --targets host + Qwen 7B if you only use one machine.
32 GB Comfortable default Full portable + Qwen2.5-Coder 7B (recommended default).
64 GB+ Room for medium tiers or multiple models Full portable + 14B/16B class models; see MODELS.md.

8 GB sticks: use --targets host

An 8 GB USB cannot hold a useful code-stick if you install the default all-target engine set:

~5–7 GB  binaries (6× Ollama + 6× opencode)
+ ~2 GB  smallest model (phi3-mini)
+ ~1 GB  headroom
≈ 8–10 GB  → over capacity before you finish

The install may start, but you will run out of space during download or model pull, or you will have almost no free space left for updates and extra models.

Do this instead on 8 GB:

code-stick install --targets host --model phi3-mini

host stages only the OS + CPU architecture you are installing from (for example windows-x64 on a typical Windows laptop). That cuts the binary footprint to roughly one Ollama tree and one opencode tree (~1–2 GB total, depending on platform) instead of six.

Trade-off: the stick only boots on that OS/arch. Plugging it into another machine (Mac, Linux, 32-bit Windows, ARM64 Surface without the ARM64 tree, etc.) will not work until you add targets:

code-stick add-targets all          # restore full portability later (needs free space)
code-stick add-targets darwin-arm64 # add one family at a time

You will get a confirmation prompt when --targets is not all (unless you pass --yes).

Examples by scenario

Smallest usable stick (8 GB, single Windows PC):

code-stick install --target E:\ --targets host --model phi3-mini --yes

Full portability on 32 GB (default product behavior):

code-stick install --target E:\ --model qwen25-coder-7b

Two OS families, not all six arches (saves space vs all):

code-stick install --targets windows,mac --model phi3-mini

Other space gotchas

Installer model picker. The UI now reserves a target-aware binary budget when greying out models — roughly ~1–2 GB for --targets host, and ~5–7 GB for the default six-target install (Windows + Linux x64 each ship ~1.6 GB of CUDA/ROCm libs; ARM64 + macOS land smaller). The picker prints the exact number for your selection. If you still want to triple-check fit on a small USB, use the cheat sheet above.

Fast install mode needs extra temporary space on the install machine (~2× model size in %TEMP% / /tmp). It does not change how much ends up on the USB. Direct uses host temp for Ollama pull scratch and deletes each engine archive from the stick right after extract. See COMMANDS.md.

FAT32 is rejected for models above ~4 GB (single-file limit). Use exFAT or NTFS. See README.md.

Adding models later — each code-stick add-model adds another full model blob under data/. code-stick remove-model frees that space.

Upgrading engines — code-stick upgrade-engine re-downloads only the targets already on the stick; it does not silently add missing OS trees.

Quick reference

Goal Command
Minimum binary footprint code-stick install --targets host
Default (fully portable) code-stick install (no --targets)
Add portability later code-stick add-targets all
Check what is on the stick code-stick status / code-stick doctor