notatestuser/ds4-control

macOS menubar app for fast local DeepSeek V4.1, with 1M context.

DS4 Control – macOS menu‑bar UI for the DeepSeek V4 model (via the ds4 server)

What it is – A native Swift menu‑bar application that lets you run, monitor, and interact with a local ds4‑server process that serves DeepSeek V4 (the Pro and Flash variants). The app does not perform inference itself; it only manages the external server, downloads the model files, and provides quick‑launch shortcuts to chat or coding agents.

Why it matters – Running a 100‑GB‑plus LLM on a Mac requires careful handling of RAM, GPU‑wired memory limits, and process supervision. DS4 Control automates those chores, shows live resource usage, and prevents you from starting a model that would exceed your machine’s Metal wired‑memory budget.

Core functionality

Feature What you get
Server lifecycle Start/stop the ds4‑server binary, watch its stderr for the listening on http://… line, poll /v1/models for health, and gracefully shut it down.
Model selector Choose V4 Pro (0813) or V4 Flash (0731). The UI checks your RAM and Metal wired‑limit and disables options that cannot fit.
Built‑in downloader Parallel, resumable download of the GGUF model files with a live progress bar. No external tools needed.
Resource widgets Mini gauges in the menu‑bar popup for unified memory, GPU, CPU, and power/ANE usage, refreshed every 2 seconds.
One‑click agents Buttons to launch a chat window, Claude Code, or Pi (local “agentic coding” tools) that talk to the running model.
1 M‑token context Configurable up to one million tokens (default for Pro; lower defaults for Flash on smaller RAM).
Startup integration Option to auto‑start the server when macOS boots.
Safety gating If the required Metal wired memory would be exceeded, the app shows the exact sysctl iogpu.wired_limit_mb=… command needed and a helper dialog.

Who should use it

  • Mac‑silicon power users who have 96 GiB + of RAM and want to run DeepSeek V4 locally.
  • Developers needing a quick way to spin up a local LLM for coding assistance (Claude Code, Pi) without fiddling with command‑line scripts.
  • Researchers who want to experiment with different reasoning‑effort levels (Standard, Instant, Max Think) and see real‑time system impact.

Installation (quick)

brew install --cask ds4-control   # Homebrew Cask provides a signed .dmg

Or download the latest signed .dmg from the GitHub releases page, or install via Homebrew Cask URL.

Building from source (for contributors)

  1. Fetch the submodule that contains the upstream ds4 server:
    git submodule update --init --recursive
  2. Patch and compile the server (required for the THINK_MAX feature):
    bash scripts/apply-ds4-patches.sh
make -C external/ds4 -j ds4-server
  3. Run the app against the built server:
    DS4_DIR="$PWD/external/ds4" swift run
  4. To create a distributable bundle:
    bash build.sh   # signs with your Apple Development certificate or falls back to ad‑hoc

How it works under the hood

  • SupervisorService – Swift Process wrapper that launches ds4‑server, parses its readiness line, polls the health endpoint, and handles graceful termination.
  • MetricsManager – Samples CPU, memory, GPU, and power via Mach, IOKit, and the private IOReport API every 2 seconds, feeding the UI widgets.
  • Feasibility – Calculates whether the selected model fits in RAM and the Metal wired‑memory limit, derives a safe default context size, and presents the user with the required sysctl command when needed.
  • All heavy‑logic (feasibility math, resumable chunk downloader, state machine) is pure Swift and covered by unit tests.

License

MIT – see the repository’s LICENSE file.

Bottom line – DS4 Control is a polished, signed macOS utility that removes the manual plumbing required to run DeepSeek V4 locally, giving you a single‑click way to start the model, monitor system health, and launch chat or coding agents.

Related

  • Project
    MiaAI-Lab/DeepSeek-v4-Flash-DSpark-2x-DGX-Spark

    A deployment recipe for serving the DeepSeek‑V4‑Flash‑Vision‑Exp model (1 M‑token context, image support) on a two‑node DGX Spark cluster using vLLM, DSpark speculative decoding, and NVIDIA‑optimized KV cache. Includes Docker images, environment‑variable configuration, optional NFS weight sharing, and many performance knobs.

  • Project
    antirez/ds4

    DwarfStar (antirez/ds4) is a native inference engine for a few open‑weight LLMs (DeepSeek V4, GLM 5.x) that runs on consumer hardware (Apple Metal, NVIDIA CUDA, AMD ROCm). It supports SSD streaming, multi‑GPU and distributed tensor‑parallel execution, provides a CLI, an interactive coding‑agent, and an HTTP server, and includes tools for model download, evaluation, and benchmarking. The project is beta‑quality and builds on the llama.cpp/GGML ecosystem.

  • Project
    thinkany-ai/dscode

    A local-first, multi-provider coding agent that defaults to DeepSeek V4 Flash and features OS sandboxing for secure, cost-aware software development.

  • Project
    raullenchai/Rapid-MLX

    Rapid‑MLX is a native Apple‑silicon inference engine that runs LLMs, diffusion, audio and video models locally and exposes a drop‑in OpenAI/Anthropic HTTP API, letting any client (LangChain, Claude Code, Aider, etc.) run fully offline on an M‑series Mac.

  • Dispatch
    Optimizing Local Inference: A Deep Dive into the DeepSeek 4 Flash Metal Engine

    An exploration of a specialized local inference engine for DeepSeek 4 Flash on Apple Metal, highlighting the shift toward hardware-specific optimization and the challenges of large-context processing.