Run DeepSeek-V4-Flash Locally: Hardware You Need (2026)

The cheapest new machine that runs DeepSeek-V4-Flash locally is a 128GB AMD Ryzen AI Max+ 395 mini PC. The Apple path is a 128GB MacBook Pro with the M5 Max. The single GPU with almost enough VRAM, the 96GB RTX PRO 6000, still cannot hold the good builds alone. This guide compares every path on memory and speed, with the memory shortage in mind.

How much memory does DeepSeek-V4-Flash need?

You need a 128GB-class machine. The smallest good build of this 284B-parameter model is the 2-bit DQ MLX conversion at a measured 96.5GB, with the UD-Q2_K_XL GGUF right behind at 96.8GB (measured quant sizes, 2026). The absolute floor is the 82.5GB UD-IQ1_S build, which trades real quality for the smaller file.

So the shopping question is simple: what is the cheapest way to put roughly 100GB of model plus KV cache headroom behind a fast memory bus? In August 2026, mid memory shortage, there are four serious answers and one trap.

What does the Mac path cost?

Disclosure: the Amazon links below are affiliate links. ModelFit earns a commission if you buy, at no extra cost to you.

A 16-inch MacBook Pro with the M5 Max chip and 128GB of unified memory is the cleanest consumer machine for this model. It is also the most expensive of the 128GB options, because Apple raised US Mac prices in June and the memory upgrade options for the M5 Max cost far more than they used to.

What you get for the money is the best measured performance of any machine on this list. The build generates about 39 tokens per second on a 128GB M5 Max in antirez's ds4 engine, and about 27 on an M3 Max, per the August 2026 README (antirez/ds4, 2026).

When this post was written you could not buy a bigger Mac instead. Apple had retired the 256GB and 512GB M3 Ultra Mac Studio tiers during the shortage, so the Mac Studio topped out at 96GB and the 128GB MacBook Pro was the largest-memory Mac sold new (why that matters for this model, 2026). Since September 22 the M5 Ultra Mac Studio sells with up to 256GB, and a 512GB model ships from late October 2026. Used 192GB M2 Ultra and 256GB M3 Ultra Studios also circulate in a thin and noisy used market, and they were the only Mac route to the higher-quality 3-bit and 4-bit builds before that.

The 96GB Mac Studio row is the trap flagged above: it is a flagship machine and cannot load even the 82.5GB floor build once macOS takes its share.

What is the cheapest 128GB machine?

The AMD Ryzen AI Max+ 395 platform, by a wide margin. A GMKtec EVO-X2 with 128GB of unified memory costs well under the MacBook Pro, though the shortage pushed it far above its early-2026 level, as the Strix Halo update explains. Check the live price before you budget. Full specs and current model picks live on our Ryzen AI Max+ 395 page.

Speed lands in the same band as the Macs. Lucebox, a vendor selling these boxes, reports the V4-Flash preview reaching up to 32 tokens per second of decode on 128GB (Lucebox, 2026). Treat that as a vendor number, but it agrees with the measured Mac figures: the 13B-active MoE design keeps decode usable on every 128GB unified-memory machine.

NVIDIA's DGX Spark is the other 128GB box, and its price has risen since launch with the memory shortage. Know what you are buying: in the ds4 measurements the Spark prefills far faster than any Mac but generates under 14 tokens per second, less than half the M5 Max rate (antirez/ds4, 2026). It is a CUDA development box first and a chat machine second.

Can a GPU rig do it for less?

No, and the flagship card cannot even do it alone. NVIDIA's RTX PRO 6000 Blackwell carries 96GB of VRAM and sells for several times the price of the AMD box. The 96.8GB Q2 GGUF does not fit in 96GB of VRAM. Only the 82.5GB quality-floor build loads fully on the card, with modest context. You would pay far more to run a worse quant.

The used route is stacking RTX 3090s. Five cards give 120GB of VRAM, but that is before the server board, the power supplies for a multi-kilowatt rig, and the assembly time. For a 13B-active MoE that already decodes at chat speed on a mini PC drawing under 200W, the multi-GPU rig is an enthusiast project, not the value play. GPU rigs earn their money on dense models and batch serving, not on this workload.

Why not a CPU server?

Because this model is too small to justify one. The used-EPYC path shines when weights outgrow every consumer machine: our Kimi K3 build guide lists a dual-socket 1TB DDR4 server part by part, and K3's smallest real quant needs 567GB, so there the server is the only local option. DeepSeek-V4-Flash needs 97GB. Every gigabyte past 128GB of capacity is wasted on it, and DDR4 bandwidth cannot match a unified-memory APU for single-user decode anyway.

The honest rule during the shortage: buy capacity for the model you actually want to run. RAM is the majority of any AI build's cost right now, and the RAM sizing guide maps exactly how much each model tier needs before you spend anything.

What if you spend nothing on hardware?

Renting or using the API is far cheaper until your usage is constant. An hourly cloud GPU rental undercuts buying the AMD box for anything but round-the-clock use. Rent first, buy when the meter proves you should.

The API case is even more lopsided. DeepSeek charges $0.14 per million input tokens on a cache miss and $0.28 per million output tokens (DeepSeek pricing, 2026). Local hardware is not how you save money on this model. You buy it for privacy, offline work, and unmetered agent loops.

Which path should you pick?

If you already own a machine, check what it runs before spending anything: the modelfit wizard covers every Mac, GPU and RAM tier, or run npx @modelfit/cli@latest in a terminal. And see how V4-Flash class open models compare with the cloud flagships on the benchmark leaderboard.

FAQ

What is the cheapest way to run DeepSeek-V4-Flash locally?

A 128GB AMD Ryzen AI Max+ 395 mini PC. It holds the 96.5GB 2-bit build with room for KV cache, and vendor-reported decode reaches up to 32 tokens per second.

Can a single GPU run DeepSeek-V4-Flash?

Not the good builds. The largest single-card VRAM pool is the RTX PRO 6000 at 96GB, and the 2-bit GGUF measures 96.8GB, so it does not fit. Only the 82.5GB 1-bit floor build loads on one card, with limited context. Multi-GPU setups work but cost more than a unified-memory box for this workload.

Is the MacBook Pro worth the premium over the AMD box?

It is the fastest option measured so far, about 39 tokens per second on the M5 Max against a vendor-claimed 32 on the AMD chip, and you get a laptop, macOS, and the MLX ecosystem with it. If the machine is your daily computer anyway, the premium is easy to defend. If it will sit on a shelf running an agent, the AMD box does the same job for much less.

Should I wait for hardware prices to drop before buying?

The 2026 memory shortage repriced every 128GB-class machine upward, and unified-memory boxes went up the most because memory is most of what you pay for. Nothing in the current supply picture points at quick relief, and our RAM price tracker post follows the story. Buy the capacity your target model needs, nothing more.

Is it cheaper to rent a GPU or use the API instead?

Almost always, at first. A 96GB cloud GPU rents by the hour, and the DeepSeek API serves this exact model at $0.28 per million output tokens. Hardware only wins once usage is constant and the machine runs day and night. Local buys you privacy and unmetered loops, not savings.

Sources

• DeepSeek API pricing: https://api-docs.deepseek.com/quick_start/pricing
• antirez, ds4 inference engine speed measurements: https://github.com/antirez/ds4
• Lucebox, DeepSeek V4 Flash on Ryzen AI MAX+ 395: https://www.lucebox.com/blog/deepseek-v4-strix-halo
• Measured DeepSeek-V4-Flash quant sizes, HuggingFace API: https://huggingface.co/api/models/mlx-community/DeepSeek-V4-Flash-2bit-DQ