DeepSeek V4 Flash Local on a Mac (2026): Inside antirez's ds4 Engine, the 96-512GB Hardware Wall, and a Buy-vs-Cloud-Rent Decision Table

Five hard gates that turn "good software" into "still cannot run it"

The reception of ds4 has been unusually warm for a one-week side project. Towards AI tested it on 18 tasks and called the experience "a lot more B than A," where B is the frontier API and A is small local models. None of that matters if you cannot get the bytes onto Metal. Five gates explain why.

1. 96 GB unified memory is the floor, not the target. The q2 GGUF is 80.8 GiB on disk. After loading, you still need headroom for KV cache and runtime. The community floor is 96 GB; 128 GB is the first comfortable rung. A 64 GB machine will swap aggressively and become unusable.
2. Metal-only path. CPU inference will crash the kernel. The README is direct: "current macOS versions have a bug in the virtual memory implementation that will crash the kernel if you try to run the CPU code." That removes any non-Apple-Silicon Mac and any hope of CPU fallback.
3. The 2-bit quant is unusual on purpose. Only routed MoE experts are aggressively quantized: IQ2_XXS for gate and up, Q2_K for down. Shared experts, attention projections, router, embed, indexer, and high-precision components stay at Q8_0, F16, or F32. The trade only works because antirez ships custom GGUFs in antirez/deepseek-v4-gguf. Generic V4 GGUFs from Hugging Face will not load.
4. On-disk KV is a feature with a hidden bill. ds4 lets the KV cache spill to a directory you choose. That makes long contexts persistent across sessions, which is rare. It also pushes serious write traffic to the SSD. Apple internal NVMe has finite TBW; running 1M-token sessions 24/7 for three years is real wear, not a thought experiment.
5. Capex is one-shot and the residual is brutal. A 128 GB MacBook Pro M3/M4/M5 Max is roughly $4,000-$5,000 in the US. A 256 GB Mac Studio M3 Ultra is around $7,000. The 512 GB top-spec Mac Studio crosses $14,000. Frontier models iterate every six to nine months. Resale value drops fast.

The honest summary: ds4 took down the software wall, but the hardware wall is taller and softer than it looks. You pay for it once at the register, then again every time the model under your engine becomes a generation old.

What ds4 actually optimizes (and why that is different from llama.cpp / MLX)

The defining bet in ds4 is single-model focus. llama.cpp targets hundreds of architectures. MLX is a general Apple Silicon ML stack. ds4 is a Metal graph executor written for the exact layer geometry of DeepSeek V4 Flash. Hand-tuned kernels, dedicated KV handling for routed MoE, on-disk KV protocol, prompt rendering, tool calling, plus an OpenAI- and Anthropic-compatible HTTP server in the same binary. The cost: when V4 Flash is superseded, much of the engine has to be rewritten. The benefit: today, on a 128 GB Mac, it is the shortest path to a working setup.

How to read this: pick ds4 if V4 Flash is the only model you care about and speed matters. Keep llama.cpp if you also rotate Llama, Qwen, Mistral, or Phi. MLX makes sense for research experiments inside the Apple ecosystem. Most teams I have seen run ds4 and llama.cpp on the same remote Mac and route requests by task.

Hardware bill: q2 / q4 / Pro mapped to real Macs

The table below combines four facts on one row: model tier, minimum unified memory, the actual Mac it implies, US retail price band, and benchmark numbers from the ds4 README. All prefill and generation figures are from the official measurement table for MacBook Pro M3 Max 128 GB and Mac Studio M3 Ultra 512 GB. US prices reflect Apple online store as of May 2026, top-spec configurations.

Why Apple Silicon is the right host (and why x86 + discrete GPU is not)

ds4 lists Metal as the primary target. That is not a marketing decision. It follows directly from the V4 Flash hardware envelope.

• Unified Memory Architecture removes the VRAM ceiling. On x86 with discrete GPUs, VRAM size is the constraint. To fit an 80-150 GB model you need an H200 (141 GB HBM3e) at five figures, or four RTX 4090s with INT4 and multi-GPU plumbing. Apple Silicon shares one large pool between CPU and GPU. A 128 GB M3 Max holds the 80.8 GB q2 GGUF, KV cache, and runtime without any PCIe round trips.
• Memory bandwidth fits the MoE pattern. V4 Flash is sparse MoE: 284B total parameters, 13B activated per token. Each token must read its routed expert weights from memory. M3 Ultra delivers around 800 GB/s of memory bandwidth. Combined with IQ2_XXS shrinking the routed experts, it stays inside the bandwidth envelope.
• Internal NVMe matches the on-disk KV protocol. Apple Silicon Macs ship NVMe SSDs in the 5-6 GB/s sequential class. The --kv-disk-dir protocol is designed to write KV slabs in patterns those drives handle well. On a typical desktop PC SATA SSD it would not be usable.

Net: a high-memory Mac is, for now, the only consumer-grade hardware that lands in the "fits, runs fast enough, costs less than a car" intersection. That is also why this wave of local frontier-model inference is happening on Mac, not on x86.

Seven steps from git clone to Cursor talking to V4 Flash

The path below is the shortest validated route from zero to "Cursor sends a prompt and DeepSeek V4 Flash answers it." It works on a local 128 GB MacBook Pro, on a local 512 GB Mac Studio, and on a remote Mac over SSH. The only difference for remote is one extra port-forwarding line.

1. Pick the tier. 96-128 GB Mac → q2. 256 GB+ → q4 (or q2 for speed). MacBook Pro caps at 128 GB so always q2.
2. Clone and build. git clone https://github.com/antirez/ds4 && cd ds4 && make. Apple Silicon defaults to Metal. Do not run make cpu; the CPU path will crash macOS.
3. Download the model. ./download_model.sh q2 for 128 GB or ./download_model.sh q4 for 256 GB+. The script pulls from huggingface.co/antirez/deepseek-v4-gguf with curl -C - resume support. q2 is 80.8 GiB; q4 is 153.3 GiB.
4. Start the server. ./ds4-server --ctx 100000 --kv-disk-dir /tmp/ds4-kv --kv-disk-space-mb 8192. Do not jump straight to --ctx 1000000; that pulls memory you may not have and triggers swap. Start at 100K and grow.
5. Verify. In another terminal: curl http://localhost:8080/v1/models. You should see deepseek-v4-flash. A POST /v1/chat/completions with a one-line prompt confirms generation works.
6. Connect Cursor or opencode. In Cursor's model settings, add a custom OpenAI-compatible endpoint. URL: http://<ds4-host>:8080/v1. Model: deepseek-v4-flash. API key: any string (ds4 does not validate by default).
7. Remote Mac via SSH local forwarding. On your laptop: ssh -L 8080:localhost:8080 user@mac-host. Then point Cursor at http://localhost:8080/v1. The experience is indistinguishable from local. Full topology in the SSH local-forward to dedicated remote Mac runbook.

Three citable data points (from the ds4 README and Hugging Face)

• Model spec. DeepSeek V4 Flash: 284B total parameters, 13B activated, 1M token context length, FP4 + FP8 mixed in the original release. This is the largest open-weight model that can plausibly run on a single high-memory Mac in 2026.
• Measured throughput. MacBook Pro M3 Max 128 GB at q2 with an 11,709-token long prompt: prefill 250.11 t/s, generation 21.47 t/s. Mac Studio M3 Ultra 512 GB at q2, same prompt: prefill 468.03 t/s, generation 27.39 t/s. q4 on the Ultra at 12,018 tokens: prefill 448.82 t/s, generation 26.62 t/s.
• GGUF on disk. q2 weighs 80.8 GiB (IQ2_XXS gate/up, Q2_K down on routed experts; Q8 on attention/shared/output; F16 on router/embed/indexer/HC/compressor; F32 on norms/sinks/bias). q4 is 153.3 GiB with Q4_K on routed experts and the rest unchanged. Optional MTP file for speculative decoding: 3.6 GiB.

Buy vs rent: a three-year TCO that pays attention to residuals

Capex stories tend to flatter the buyer because they rarely model frontier-model iteration risk or SSD wear. The table below does. Buy side assumes US retail and a 50% three-year residual, which is generous for hardware that will be a generation behind by 2027. Cloud-rent side uses MACCOME's public rate card and assumes flexible monthly use, with hourly available for short experiments. Power, networking, replacement, and physical space are excluded from the buy side; including them widens the gap.

Reframed in plain language: buying the 512 GB Ultra is paying $14,500 today plus three years of iteration risk plus SSD wear. Renting moves all three of those risks to the platform. The math only favors buying if you are sure you will saturate the box for 18+ months. The same logic shows up on a smaller dollar scale in the Mac mini M4 buy-vs-rent decision matrix, except here every variable is multiplied by ten and the residual risk is sharper.

Closing the loop: you do not need to spend $14k to run V4 Flash

antirez has demonstrated something specific. Running a 284B-parameter frontier model locally is now a software-solved problem. 18,404 lines of C, Metal-first, on-disk KV, OpenAI- and Anthropic-compatible endpoints, an integrated coding agent. The barrier left is the 96 to 512 GB hardware wall, and for individuals it is not just expensive: it is brittle.

If you actually account for the full picture, owning a top-spec Mac Studio Ultra to run ds4 has three quiet costs. (a) A $4,000 to $14,500 one-time hit that locks cash flow and depreciates fast as the next model lands. (b) Real SSD wear from running 1M-context sessions over years. (c) Zero elasticity: today you want q2, next quarter you want q4, but you still own only one machine. For anyone running a serious AI Agent stack or a small team that wants to share a single Ultra across experiments, MACCOME's high-memory remote Macs (128 / 256 / 512 GB tiers) on monthly or hourly rent are usually the cleaner answer. The Capex becomes Opex, the residual risk moves to the platform, and the same machine can be split among teammates.

To plan node selection by region, see the 2026 multi-region Mac node guide. To pair local inference with cloud-API routing, the 2026 OpenRouter rankings and routing matrix covers the "cloud API + local inference" layer that complements ds4. Read in that order, the picture is complete: where to host the model, and how to layer it with managed APIs.