Salvatore Sanfilippo (antirez), author of Redis, open-sourced ds4 (DwarfStar 4) in 2026: a local inference engine built only for DeepSeek V4 Flash and PRO, implemented in plain C, not wrapped around llama.cpp and not aimed at the general GGUF market. The README states the goal plainly: make local inference on a top-tier personal machine or Mac Studio trustworthy enough to replace everyday Claude or GPT calls, with official vector checks, long-context tests, and coding-agent integration.

Within days the project passed 11,000+ GitHub stars. Hacker News and community reviews converged on one point: a 284B-class MoE running offline on a MacBook with tool calling and six-figure context windows. That is a different conversation from the 2025 era of 7B toy models. ds4 moved the debate from "can it run" to "would you ship production code on it." For KVMNODE customers the hype itself signals a steepening demand curve for high-memory Macs; what blocks most teams is the memory wall in the next section.

The ds4 story compresses to max out Apple Silicon, push long session state to SSD. Reported capabilities from the project and early reviewers include the following.

On a MacBook Pro M5 Max, ds4 has been quoted at roughly 463 token/s prefill and about 34 token/s generation (quantization and context length move the numbers). That places the engine in the first tier among consumer hardware. It advertises a context window up to about one million tokens, paired with DeepSeek V4's compressed KV design, so "entire repo plus long chat" becomes a planning scenario instead of a demo slide.

On-disk KV persistence is a differentiator: session KV can be written to fast Mac SSDs so restarts or task switches avoid a full prefill replay. That matters for laptops that sleep daily and for agents that must resume yesterday's thread. 2-bit asymmetric quantization compresses routing experts aggressively while keeping other layers precise, which is how Flash becomes viable on 128GB machines. ds4-server exposes OpenAI- and Anthropic-compatible endpoints, so Cursor, opencode, and Claude Code can treat the instance as a private model vendor.

git clone https://github.com/antirez/ds4
cd ds4 && make
./ds4-server --ctx 100000 --host 127.0.0.1 --port 8080

The README also warns that on macOS the CPU inference path can trigger kernel virtual-memory bugs; production should use Metal (or CUDA on Linux). That belongs on your cloud Mac runbook alongside the health probes in the diagnostic ladder.

No matter how elegant the engine, unified memory capacity sets the ceiling. The table below merges official README guidance, community measurements, and typical retail pricing (USD approximations; configs and FX move the numbers). Use it for budget or rental decisions: separate "can run" from "runs comfortably."

For teams the pragmatic split is validate ds4 separately from daily iOS CI: 16GB·256 or 24GB·512 for builds and OpenClaw, 128GB+ dedicated to ds4-server, so DerivedData and million-token KV never fight on one SKU. Sizing detail lives in storage and memory pairing.

Listing Metal as the primary macOS backend is not marketing. Apple Silicon's unified memory architecture (UMA) lets CPU, GPU, and Neural Engine share one physical pool, avoiding the PC split where "24GB VRAM plus 64GB RAM" caps what you can load. For large-model inference, one addressable space directly bounds quantized weights and KV size. M3/M4/M5 memory bandwidth pushes prefill throughput to the consumer ceiling.

macOS NVMe pairs with ds4's disk KV as a second coupling: long sessions need not sit entirely in RAM; cold starts can reload context blocks from SSD. A Linux plus CUDA path (including DGX Spark tuning) exists in the repository, but for developers who already own Macs and want offline coding, high-memory Mac is the best consumer platform for ds4 today, consistent with antirez's Hacker News comments.

Running ds4 in macOS VMs on non-Apple hardware or hackintosh setups sacrifices Metal stability and licensing. Cloud deployments should use real bare-metal Apple Silicon nodes, which is why KVMNODE delivers dedicated Mac Mini hardware rather than "Mac-like" virtual desktops.

The steps below assume you ordered a cloud Mac with 128GB or more unified memory. Pick region by Git remote and weight download source per six-region guidance. Model files are large; align with object storage or a Hugging Face mirror in the same region to avoid trans-Pacific tail latency.

For technical reviews or procurement memos, cite these public figures (they move with upstream README updates): ① GitHub 11k+ stars reflecting May 2026 community momentum; ② community-reported ~463 t/s prefill and ~34 t/s generation on MacBook Pro M5 Max (quantization and context dependent); ③ official production floor at 96GB unified memory, with 128GB as the safer Flash long-context tier.

Lay alternatives side by side. Staying on Claude / GPT APIs only bills per token and sends code plus long context over the internet, which hurts IP-sensitive work. Buying a Mac Studio Ultra locks CapEx in the tens of thousands and freezes upgrade cycles. Forcing ds4 onto generic Linux GPU clouds leaves Metal optimizations unused and changes MoE memory topology requirements. Renting 128GB / 512GB cloud Macs from KVMNODE by the hour or month turns ds4's top-tier local inference into switchable OpEx: real Metal hardware, team sharing, and data that stays on a dedicated instance—ideal to validate before hardware purchase.

Teams running iOS CI, OpenClaw Gateway, and ds4 together should split pools physically or logically; do not merge 16GB build SKUs with 128GB inference SKUs. Tiers and pricing on the pricing page, ordering on the order page, runbooks in the help center.