pi-ds4

Run antirez/ds4 models locally in Pi. The extension supports DeepSeek V4 Flash, DeepSeek V4.1 Flash, DeepSeek V4 Pro, and GLM 5.2.

Install

pi install https://github.com/mitsuhiko/pi-ds4

Restart Pi or run /reload.

Get started

1. Run /ds4 and choose Download model.
2. Pick a model. DeepSeek V4 Flash Q2 is the smallest option and needs about 96 GB of RAM.
3. After the download finishes, run /model and select the new ds4/... model.
4. Start chatting.

Downloads report a progress bar with transfer rate and ETA in the status line, counted from the bytes on disk, so resuming an interrupted download picks up where it left off. Full output stays in ~/.pi/ds4/log.

Models are never downloaded automatically. Once installed, they are discovered on future Pi starts and shown in /model. Selecting one starts ds4-server on demand; it stops automatically when no Pi processes are using it.

Use /ds4 at any time to download another model, view the log, or start and stop the server manually.

Models

These are conservative full-residency recommendations. ds4 can also stream from SSD or run across multiple machines, with different performance tradeoffs. Signing in to Hugging Face before downloading may improve download speed.

DeepSeek V4.1 Flash needs much more disk than RAM: 341 GiB on disk for Q2 and 483 GiB for Q4, because its Engram tables stay on disk and are read as needed. Only 152 GiB (Q2) and 294 GiB (Q4) of weights are resident. Keep the GGUF on a fast local SSD; the Q4 download arrives in two parts that are joined automatically and needs another 37 GiB of free space while joining. V4.1 runs on Metal, and on CUDA for text inference. Its GPU duty cycle cannot be throttled, so power is ignored for it.

Images

DeepSeek V4.1 Flash can read images. Run /ds4 → Download model and pick the vision encoder (about 0.9 GiB); the next server start adds --vision and the V4.1 models accept image input in Pi. Upstream only supports the V4.1 encoder on Metal, so the option is hidden elsewhere. ds4-server accepts up to 16 inline images per request.

SSD streaming

When a model's resident weights do not fit in system RAM, the extension starts ds4-server with --ssd-streaming and caps the context at 131072 tokens unless contextTokens is set explicitly. Today only the V4.1 models declare a resident footprint, so this affects them alone. Set ssdStreaming to true or false to decide it yourself.

Configuration

Most users do not need any configuration. Optional settings live in ~/.pi/ds4/settings.json:

{
  "$schema": "https://raw.githubusercontent.com/mitsuhiko/pi-ds4/main/settings.schema.json",
  "protocol": "openai-responses",
  "readyTimeoutMs": 900000
}

See settings.example.json for all options. Each setting also has a DS4_* environment variable; environment variables take precedence. Restart Pi or run /reload after changing settings.

Set power to an integer from 1 to 100 to limit the GPU duty cycle for DeepSeek models, reducing heat and fan noise. It defaults to 100. GLM 5.2 and DeepSeek V4.1 always run at full power because upstream does not support throttling them.

Runtime files, downloaded models, caches, and logs are stored in ~/.pi/ds4. Package-managed ds4 checkouts update automatically; explicitly configured or local development checkouts are left untouched.

Local development

To use this extension checkout with an existing ds4 checkout:

./install-pi-extension-local.sh /path/to/antirez-ds4-checkout

If ~/.pi/ds4/support already points elsewhere, rerun with --force. Existing model downloads are preserved when possible.