DS4 Control
A macOS menu bar pane for DeepSeek V4 and V4.1 via dwarfstar.
It launches, supervises, and monitors a local ds4 server, lets you pick V4.1 Flash 1017, V4 Pro (0813), or V4 Flash (0731) with up to 1M context, and shows resource use.
Launch Pi, Claude Code or BYOC (Bring Your Own CLI) for local agentic coding.
-1f6feb%3Fstyle%3Dfor-the-badge%26logo%3Dapple%26logoColor%3Dwhite)
Signed with a live Apple Developer ID & notarized by Apple
Quick Install
brew install --cask ds4-control
Features
- Start / stop / monitor the local
ds4-server child process — spawn, stderr readiness detection, health polling, graceful stop, and crash detection.
- Pro / Flash selector with a RAM feasibility checks.
- Model downloads via a built-in native parallel downloader, with a live progress bar and resume across restarts.
- Mini resource widgets: unified memory, GPU, power, and CPU, sampled on a timer.
- Launch Chat to talk to the model.
- Launch Claude Code or Pi to plan, write, maintain or refactor code.
- 1M Context configurable in settings.
- Launch on macOS Startup
What it is not:
- No model search or registry browsing.
- No embedded inference — all inference is delegated to
ds4-server.
Dev quick start
antirez/ds4 is vendored as a git submodule at external/ds4:
git submodule update --init --recursive # fetch ds4 into external/ds4
bash scripts/apply-ds4-patches.sh # THINK_MAX prefix (antirez/ds4#635)
make -C external/ds4 -j ds4-server # build the ds4-server binary
DS4_DIR="$PWD/external/ds4" swift run # build + run the dev app against the submodule
Requirements
- Apple Silicon
- You don’t pre-download the model — DS4 Control downloads it for you with a built-in parallel downloader, resumable across restarts.
- Auth (optional): the model repository is public, so no token is required for normal use.
- RAM — see below.
RAM feasibility
DeepSeek V4 is memory-hungry so DS4 Control gates feasibility before launching.
| Variant |
Quant |
RAM |
Notes |
| V4 Pro (0813) |
pro-imatrix |
≥ 512 GiB required |
Anything below is blocked. |
| V4.1 Flash 1017 |
1017-q2 |
≥ 96 GiB |
341 GiB on disk, ~152 GiB resident main weights + ~189 GiB disk-only Engram. SSD streaming engages automatically when the resident configuration does not fit (usually 96–255 GiB; slow: routed experts stream from disk). |
| V4.1 Flash 1017 |
1017-mxfp4 |
≥ 256 GiB |
467 GiB on disk, ~278 GiB resident main weights + ~189 GiB disk-only Engram. Streams on lower tiers; fully resident when the context, sessions, and live Metal wired limit permit it. |
| V4 Flash (0731) |
q4-imatrix |
≥ 256 GiB |
Standard. |
| V4 Flash (0731) |
q2-imatrix |
96 GiB minimum |
96–127 GiB requires raising the Metal wired limit (see in-app help). |
V4.1 defaults to Q2 below 512 GiB and MXFP4 from 512 GiB. Its automatic context is 32,768 below the full-residency RAM tiers (Q2 below 256 GiB and MXFP4 below 384 GiB), and 1,048,576 at or above them. Whether a launch uses SSD streaming also depends on the live Metal wired limit. The launch gate accounts for each configured chat/agent session and ds4’s host-memory budget, so the minimum tier alone does not guarantee every context/session combination fits. Engram stays on disk in both modes.
V4.1 Flash 1017 downloads and migration
The 1017 recipes are pinned to Hugging Face revision 9dcee516: Q2 downloads one verified GGUF; MXFP4 downloads nine standard GGUF shards and assembles them natively into one file. The shards contain independent headers and tensor offsets, so they are parsed and rewritten rather than concatenated. Each shard is SHA-256 verified and removed only after its tensors and a durable assembly checkpoint are saved. Interrupted downloads and assembly resume after restart. MXFP4’s incremental assembly needs about 562 GiB peak disk space plus 1 GiB headroom; the finished file uses about 467 GiB. No Python installation is needed.
An existing V4.1 Q2 choice stays Q2; the former V4.1 Q4 choice becomes 1017 MXFP4. Other model selections and server/chat preferences are preserved. Earlier V4.1 weights are never renamed or treated as 1017 downloads, and the new weights use separate, quant-specific KV cache directories.
The popup detects obsolete V4.1 GGUFs, download sidecars, and the old kv-flash-41 cache and offers a cleanup review while the server is stopped. A native confirmation lists the exact paths and estimated space before anything is deleted; Cancel is the default. Not now hides that banner, and Settings → Downloads → Clean up old V4.1 data… remains available. Current 1017 files and unrelated Hugging Face cache entries are kept. Failed deletions remain visible for retry.
Measured single-stream generation throughput on a Mac Studio M3 Ultra (512 GiB):
| Model |
Throughput |
| V4 Pro (0813) |
~14 tok/s |
| V4 Flash (0731) |
~35 tok/s |
Varies with context length, prompt, and the Metal wired limit.
Coding Agents
It’s suggested to follow the ds4 coding agent setup guide to configure your OpenCode/Claude/Codex/Pi as you prefer.
Build & Run
For development:
swift run
To produce a distributable bundle:
bash build.sh
This builds a release binary and assembles DS4 Control.app.
The bundled runtime is pinned to antirez/ds4 a1012463383a. The app’s V4.1 allocation/admission mirror is checked against that source. The manual memory harness can exercise either 1017 quant in streaming or resident mode; it is separate from the unit tests and does load the real weights:
# Q2 streaming frontier run on a 128 GiB host.
DS41_LIMIT_GIB=128 scripts/flash41-mem-harness.sh "32768"
# Short resident MXFP4 smoke run on a suitably sized host.
DS41_QUANT=mxfp4 DS41_MODE=resident DS41_LIMIT_GIB=512 \
DS41_PROMPT_TOKENS=64 DS41_SKIP_CONTROL=1 scripts/flash41-mem-harness.sh "32768"
These commands are validation procedures, not a claim that this release has been tested with real 1017 weights on every memory tier. A short smoke prefill does not validate the frontier-memory peak.
First run: open Settings (the gear in the popup) and set the ds4 directory — the folder that contains ds4-server.
Signing
build.sh auto-detects your Apple Development identity via security find-identity and signs the bundle with it. If no Apple Development identity is installed, it falls back to ad-hoc signing (the app runs locally but is not distributable).
To sign with your own key:
- Install an Apple Development certificate (Xcode → Settings → Accounts → Manage Certificates → + → Apple Development), or
- Set
DS4_SIGN_IDENTITY="Apple Development: …" before running build.sh.
How it works
DS4 Control is a single Swift binary — no embedded inference and no second process language. Three @MainActor objects do the work, and the SwiftUI layer just observes them:
SupervisorService owns the ds4-server lifecycle through Foundation.Process: it builds the launch arguments, watches stderr for the listening on http:// readiness line, polls GET /v1/models for health, and stops gracefully with SIGTERM (SIGKILL fallback). Model weights are fetched by a built-in native parallel downloader (resumable across restarts).MetricsManager samples CPU, memory, GPU, and power/ANE via Mach, IOKit, and the private IOReport interface every 2 s, publishing a SystemSnapshot to the widgets.Feasibility turns installed RAM into a variant choice and a budget-derived default context (pure, fully unit-tested).
Attribution
- A big thanks of course to @antirez for antirez/ds4, llama.c and GGML for instrumental foundational work.
- The resource collectors and widgets are adapted from mac-resource-monitor, which in turn credits macmon (MIT) for the IOReport power-sampling approach.
- The server-supervision pattern is built on the lineage of mlx-serve.
License
MIT — see LICENSE.
评论