The honest answer is that DeepSeek V4 has no VRAM requirements, because you cannot run it locally. DeepSeek has not released V4 or V4 Pro weights, and the parameter count is undisclosed. It is a cloud API model. Any page quoting a VRAM figure for DeepSeek V4 is extrapolating from a model whose size is not public. We checked the DeepSeek site and their Hugging Face org on September 3, 2026: no V4 weights exist. This guide covers what is actually known, and what to run locally if you want DeepSeek-class output.
TL;DR: DeepSeek V4 and V4 Pro are cloud-only. No weights, no GGUF, no local path. If you want a local reasoning model in the DeepSeek lineage, DeepSeek-R1 Distill Llama 70B (42 GB) runs on 64GB+ Macs, and the Qwen 3.5 122B-A10B class is the current local frontier. Do not trust VRAM figures for a model with undisclosed size.
Why is there no local DeepSeek V4?
DeepSeek's last open-weight flagship was the R1 / V3 generation, and V4 shipped as an API product instead. The size is undisclosed, so even the memory math is impossible. A VRAM figure requires a parameter count. DeepSeek has not published one for V4. The 671B figure people cite belongs to DeepSeek-R1, a different, older, open model. In fact, that mix-up is where most of the invented VRAM numbers come from.
This matters because the local-AI ecosystem now has two kinds of frontier models. The open-weight ones can be sized for hardware (Qwen 3.5 122B, Llama 4, GPT-OSS 120B). The closed API ones cannot (DeepSeek V4, the GPT and Claude lines). A sizing guide only means something for the first kind. Every honest VRAM figure starts from a published parameter count and a real file you can download and weigh. Guides that blur that line are guessing, and the guesses spread fast. Once one site invents a number, a dozen others cite it. Would you buy hardware on a guessed parameter count? Neither would we.
Which DeepSeek models can you actually run?
The DeepSeek lineage that runs locally is the R1 distillation family, from the open era:
| Model | Size | Loaded | Min machine |
|---|---|---|---|
| DeepSeek-R1 Distill Qwen 7B | 7B dense | ~5.5 GB | 12 GB |
| DeepSeek-R1 Distill Qwen 14B | 14B dense | ~11 GB | 16-24 GB |
| DeepSeek-R1 Distill Llama 70B | 70B dense | ~42 GB | 64 GB+ |
| DeepSeek-R1 full 671B | 671B MoE | ~380 GB | 512GB Mac Studio |
The 70B distill is the practical ceiling for most people. At 42 GB it fits a 64GB Mac Studio and runs at roughly 10-14 tok/s est. — full detail on the R1 Distill 70B model page. The full 671B, however, is a 512GB Mac Studio M5 Ultra model and even there it is slow. These are reasoning models from early 2025. They are competent, but the 2026 Qwen and GPT-OSS releases have moved past them at lower memory.
What does the local frontier look like in 2026?
If the goal is DeepSeek-V4-class reasoning on your own hardware, these are the models that actually ship weights:
- Qwen 3.5 122B-A10B (~72 GB) — the strongest open reasoning model we track. 96GB machines. See the 122B requirements guide.
- GPT-OSS 120B (~65 GB) — OpenAI's open MoE, lighter than the Qwen at the same tier, 96GB machines. See its best-hardware page.
- Qwen 3.6 35B-A3B (~22 GB) — the 32GB-tier MoE that delivers most of the agentic reasoning at a third of the memory. See the 35B-A3B model page.
- GPT-OSS 20B (~14 GB) — for 24GB machines, the best reasoning per GB available locally. See the GPT-OSS 20B model page.
None of these is DeepSeek V4, and none of them pretends to be. All of them are real, downloadable, and sized from published specs rather than guessed. You can verify every figure in this guide yourself before spending a dollar. From our analysis of the catalog, that distinction — published versus guessed — is now the main quality signal in this niche. It is why this guide ends with alternatives instead of a number.
When should you use the API instead?
If you specifically need DeepSeek V4, the answer is the API, not local hardware. The local case for it does not exist. That is a legitimate choice. Cloud APIs make sense for occasional frontier-class queries. Local models make sense for privacy, cost at volume, and offline work. The mistake is buying hardware for a model that has no local build. Sizing pages that never checked whether weights exist encourage it. Consider it a rule: no weights, no VRAM number.
How do you verify a sizing claim yourself?
The check takes less than a minute and works for any model, not just this one. First, look for the weights. If a model is genuinely local, there is a public repository with files you can download. The deepseek-ai Hugging Face org shows exactly what DeepSeek has and has not released. Second, look for the parameter count in an official source rather than a roundup. Every honest VRAM figure is arithmetic on top of that number. Third, run the math yourself: weights plus context against your fast memory. Or let our hardware fit checker do it with published specs.
What you will find, applying that check to DeepSeek V4 today, is a blank at step one. That is not a gap in our data; it is the answer. A page that skips the check to hand you a tidy number is telling you something about its process, not about the model.
FAQ
Can you run DeepSeek V4 locally?
No — DeepSeek has not released V4 weights, and the model is available only through the DeepSeek API. The most capable local DeepSeek-lineage model is the R1 Distill Llama 70B, which needs a 64GB+ Mac.
What is the closest open-weight model to DeepSeek V4?
Qwen 3.5 122B-A10B is the closest in capability among models with public weights. It runs on 96GB machines at ~27-41 tok/s est. GPT-OSS 120B is the lighter alternative at the same tier.
How much VRAM for DeepSeek R1?
The full R1 671B needs about 380 GB, so a 512GB Mac Studio M5 Ultra. The R1 Distill 70B needs 42 GB, so 64GB or more. The 14B and 7B distills fit 16-24GB machines comfortably.
Why do some sites list a VRAM number for DeepSeek V4?
They are extrapolating from a guessed parameter count, and since DeepSeek has not disclosed V4's size, any specific figure they print is invented. We would rather tell you it is cloud-only than hand you a number we cannot defend.
___
The local alternatives in this guide are sized by the ModelFit engine from published specs, checked on September 3, 2026, with speeds labeled est. How we work: about ModelFit. Corrections welcome: contact.Match this model to a machine that can run it: by RAM tier for Apple Silicon, or by VRAM for an NVIDIA GPU.
The weekly local-AI refresh
New open-weight models, real Apple Silicon benchmarks, and the one model worth running on your Mac this week. Free, one email a week, unsubscribe anytime.
By subscribing you agree to our Privacy Policy and to receive the weekly email. Unsubscribe anytime.
Have questions? Reach out on X/Twitter