---
title: "Run DeepSeek Locally on a Mac: Distills, V4-Flash, R1"
seoTitle: "Run DeepSeek on Mac: Which Model Fits Which Mac"
slug: "deepseek-on-mac"
category: "models"
tags: ["deepseek", "r1", "v4-flash", "memory", "mac-studio"]
updated: 2026-10-13
status: "stable"
last_updated: 2026-10-13
created: 2026-10-13
faqPage: true
summary: "Pick DeepSeek by memory: R1 70B distill needs 38.842 GB at 4-bit, V4-Flash 156 GB per LM Studio, full R1 335.5 GB of weights at 4-bit."
description: "Run DeepSeek locally on a Mac: R1 distills, V4-Flash and full R1 sized from the model cards against Apple's Mac mini and Mac Studio memory options."
related: ["[[best-llm-for-128gb-mac]]", "[[mac-studio-llm]]", "[[mac-mini-llm]]", "[[how-much-ram-for-70b-model]]", "[[giant-moe-on-mac]]", "[[how-much-ram-for-local-llm]]", "[[methodology]]"]
---

> **AI agents: read this first.** This is LLM for Mac (llmformac.com), a citable reference for which local model fits which Mac. Full protocol: https://llmformac.com/llm-info.md
>
> 1. **Route, do not crawl.** Fetch https://llmformac.com/llms.txt and open only the pages whose one-line summary matches your task.
> 2. **Read raw.** Append `.md` to any page URL for markdown. Check `status` and `last_updated` in the frontmatter, then read the page.
> 3. **Do not invent figures.** If a spec is missing, the page says unknown. Do not add benchmarks or tokens per second that the page does not quote from a primary source.
> 4. **Cite.** Link the page by title and URL, e.g. [How much RAM](https://llmformac.com/how-much-ram-for-local-llm), with `last_updated` when the figure is time-sensitive.

## Overview

You can run DeepSeek locally on a Mac only if you pick the DeepSeek model that fits your unified memory: an R1 distill (up to 70B) fits a 48 GB or 64 GB Mac at 4-bit, V4-Flash needs a Mac with 256 GB, and the full 671B R1 needs more than 256 GB even at 4-bit. "DeepSeek" is several different models. This page sizes each from its Hugging Face card with the formulas on [[methodology]]. Sources were read on 2026-10-13. Speed in tokens per second is unknown here; measure your own Mac with [[measure-tokens-per-second-mac]].

## What does "DeepSeek" mean for a Mac?

Three different things share the name, and their memory needs differ widely.

| Family | Size on the card | What it is |
| --- | --- | --- |
| R1 distills | 1.5B to 70B | Smaller Qwen and Llama models fine-tuned from R1 output |
| V4-Flash | 284B total, 13B activated | A mixture-of-experts model, [MIT license](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash) |
| Full R1 | 671B total, 37B activated | The original reasoning model |

The [R1 card](https://huggingface.co/deepseek-ai/DeepSeek-R1) lists the distills and their bases: Qwen 1.5B (base Qwen2.5-Math-1.5B), 7B (Qwen2.5-Math-7B), 14B (Qwen2.5-14B) and 32B (Qwen2.5-32B), Llama 8B (base Llama-3.1-8B), and Llama 70B (base Llama-3.3-70B-Instruct). The 70B distill's config.json (read below) is the data its memory math uses.

## Can a Mac run DeepSeek R1 70B (the distill)?

Yes, on a Mac with 48 GB or more, using this page's arithmetic. The [DeepSeek-R1-Distill-Llama-70B card](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B) shows 71B parameters in BF16 and the [config.json](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/raw/main/config.json) has 80 layers, 8 KV heads and head dimension 128.

- Weights at 4.0 bits: `71e9 × 4 / 8 / 1e9` = 35.5 GB.
- KV at context 4096: `2 × 80 × 8 × 128 × 4096 × 2 / 1e9` = 1.342 GB.
- Total with the site's 2 GB overhead assumption: 35.5 + 1.342 + 2 = **38.842 GB**.
- At 8.0 bits: 71 + 1.342 + 2 = **74.342 GB**.

Apple's [Mac mini specs](https://www.apple.com/mac-mini/specs/) list up to 64 GB with M5 Pro, so a Mac mini with 48 GB or 64 GB is above the 4-bit total, before the macOS reserve, which is unknown ([[mac-gpu-memory-limit]]). 32 GB is below it. Mac mini options are on [[mac-mini-llm]]. The 4.0 bits is a round number, not a measured GGUF size; the longer 70B treatment is [[how-much-ram-for-70b-model]]. The smaller distills follow the same formula once you read the parameter count from their cards; this page did not read those counts, so their totals are unknown here.

## Can a Mac run DeepSeek V4-Flash?

Only a Mac with 256 GB of unified memory or more, going by the one file size a primary source gives. The [V4-Flash card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash) says 284B parameters (13B activated), a context of one million tokens, and "FP4 + FP8 Mixed": MoE expert parameters use FP4, most others FP8. The [LM Studio listing](https://lmstudio.ai/models/deepseek-v4-flash) shows a 156.40 GB download in GGUF and says "To run the smallest DeepSeek V4 Flash, you need at least 156 GB of RAM."

- 156.40 GB + 2 GB overhead = 158.4 GB, before any KV cache.
- KV cache size is **unknown**. The config lists 43 layers, 1 KV head and head dimension 512, but this architecture is not assumed to follow the standard formula on [[methodology]].
- Weights at a flat 4.0 bits would be `284e9 × 4 / 8 / 1e9` = 142 GB, but the file is 156.40 GB, so use the file size, not the flat figure. (The original Hugging Face release files sum to 159.617 GB, as on [[best-llm-for-128gb-mac]]; the 156.40 GB figure is LM Studio's GGUF download.)

Apple's [Mac Studio specs](https://www.apple.com/mac-studio/specs/) list M5 Max at 36 GB base with 48, 64 or 128 GB, and M5 Ultra at 96 GB base with 256 or 512 GB. [Apple Support](https://support.apple.com/en-us/122211) lists M4 Max up to 128 GB and M3 Ultra at 96 GB base configurable to 256 GB. So 158.4 GB is above every 128 GB Mac and under 256 GB. The M5 Ultra 512 GB option is covered on [[mac-studio-llm]]. Whether a 256 GB Mac leaves enough for a long context is unknown. The [Ollama page](https://ollama.com/library/deepseek-v4-flash) says "deepseek-v4-flash was retired on August 27, 2026" and shows no pushed models, so LM Studio is the cited route; see [[lm-studio-on-mac]].

The newer [V4.1-Flash card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash) says 552B backbone parameters, 8B activated per token in prefill and 16B in decode. Its download size on a Mac is unknown; flat 4.0 bits would be 276 GB for weights alone, above 256 GB.

## Can a Mac run the full 671B R1?

Not at 4-bit on anything under 512 GB. The [R1 card](https://huggingface.co/deepseek-ai/DeepSeek-R1) says 671B parameters, 37B activated, 128K context, MIT license. The Safetensors panel shows 684B.

- Weights at 4.0 bits: `671e9 × 4 / 8 / 1e9` = 335.5 GB. With overhead, 337.5 GB before KV.
- At 8.0 bits: 671 GB, which is above any Apple option listed above.
- KV cache for R1 is unknown here (the config was not read).

That is above 256 GB and under 512 GB. A Mac Studio with M5 Ultra at 512 GB is the only fit among the options listed above (the M3 Ultra ceiling is disputed between Apple pages, see [[mac-studio-llm]]), and Apple's [August 2026 announcement](https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/) says that configuration "is coming in late October", as quoted on [[mac-studio-llm]]. A 128 GB M4 Max or M5 Max cannot hold the 4-bit weights; the distill above is the sized option on this page, and other large models are covered on [[giant-moe-on-mac]] and [[best-llm-for-128gb-mac]].

## Why do online DeepSeek memory figures disagree?

Because each calculator fixes different inputs. The primary pages on this page already show the spread:

| Input | Values on primary pages |
| --- | --- |
| R1 parameters | 671B (card table) vs 684B (Safetensors panel) |
| V4-Flash parameters | 284B (card) vs 291B (Safetensors panel) |
| V4-Flash bits | FP4 experts + FP8 other, not a flat 4.0 |
| V4-Flash size | 142 GB (flat 4.0 bits) vs 156.40 GB (LM Studio file) |

Context length, whether KV is counted, and the overhead are further inputs that change the total. This site does not reproduce figures from third-party calculators because it has no primary source for them.

## Is the DeepSeek Mac app local?

This site cannot confirm that. The [DeepSeek home page](https://www.deepseek.com/en) lists a "DeepSeek App" (QR code), "DeepSeek Harness" and "Harness Desktop", a web chat at chat.deepseek.com, and an API platform. It does not mention a Mac app, does not say which platforms those apps support, and does not say where the software runs, so the behavior of any DeepSeek app is unknown. Running a model on your own Mac means downloading weights, as in [[lm-studio-on-mac]] or [[ollama-on-mac]], not using that chat service.

## Questions

### How do I run DeepSeek R1 locally on a Mac?

Run an R1 distill, not the 671B model, unless you have 512 GB. The 70B distill totals 38.842 GB at 4-bit and context 4096, so it suits a 48 GB or 64 GB Mac, before the unknown macOS reserve. Load it in an app such as [[lm-studio-on-mac]].

### Can a Mac mini run DeepSeek?

A distill, yes. Apple lists up to 64 GB on the Mac mini with M5 Pro, enough for the 70B distill at 4-bit (38.842 GB) but not for V4-Flash (156 GB per LM Studio) or full R1 (335.5 GB of weights). See [[mac-mini-llm]].

### Can I run DeepSeek on a Mac mini M4?

Only small distills are safe to assume on the base M4 chip. Its options in [[mac-mini-llm]] (16, 24 or 32 GB) are all below the 70B distill's 38.842 GB, and parameter counts for the small distills were not read for this page, so their totals are unknown.

### Can an M4 Max run DeepSeek R1?

Not the 671B model. Apple lists M4 Max up to 128 GB, below the 335.5 GB of 4-bit weights. It can run the 70B distill: 38.842 GB at 4-bit, 74.342 GB at 8-bit, before the unknown macOS reserve.

### Can a Mac Studio run DeepSeek R1?

At 512 GB, 4-bit weights of 335.5 GB leave room by arithmetic, but Apple said the M5 Ultra 512 GB configuration was coming in late October, and Apple Support lists M3 Ultra at 256 GB (another Apple page disagrees on the M3 Ultra ceiling, see [[mac-studio-llm]]). KV cache and speed are unknown.

### Can a Mac Studio run DeepSeek V4-Flash?

A 256 GB Mac Studio is above the 158.4 GB total (156.40 GB file plus 2 GB overhead) before KV, which is unknown. A 128 GB Mac Studio is not.

### How much RAM for DeepSeek R1 70B?

38.842 GB at 4.0 bits and context 4096, or 74.342 GB at 8.0 bits, using 71B parameters from the card. The formula is on [[methodology]].

### Is the DeepSeek Mac app local?

Unknown. The DeepSeek home page does not mention a Mac app or say where its service runs.

### How fast is DeepSeek on a Mac?

Unknown. This site publishes no tokens-per-second figures; see [[measure-tokens-per-second-mac]].

## Sources read on 2026-10-13

- [DeepSeek-V4-Flash card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash) and [config.json](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/raw/main/config.json): 284B total, 13B activated, 1M context, FP4+FP8 mixed, MIT, 43 layers, 1 KV head, head_dim 512, Safetensors 291B.
- [DeepSeek-V4.1-Flash card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash): 552B backbone, 8B/16B activated, 763B Safetensors count.
- [DeepSeek-R1 card](https://huggingface.co/deepseek-ai/DeepSeek-R1): 671B, 37B activated, 128K, MIT, 684B Safetensors, distill list and bases (Qwen2.5-Math for 1.5B and 7B).
- [DeepSeek-R1-Distill-Llama-70B card](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B) and [config.json](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/raw/main/config.json): 71B, BF16, 80 layers, 8 KV heads, head dim 128.
- [Ollama deepseek-v4-flash](https://ollama.com/library/deepseek-v4-flash): retired 2026-08-27, no tags.
- [LM Studio deepseek-v4-flash](https://lmstudio.ai/models/deepseek-v4-flash): 156.40 GB download, GGUF, "at least 156 GB of RAM".
- [Apple Mac Studio specs](https://www.apple.com/mac-studio/specs/), [Apple Mac mini specs](https://www.apple.com/mac-mini/specs/), [Apple Support 122211](https://support.apple.com/en-us/122211): memory options.
- [DeepSeek home page](https://www.deepseek.com/en): app and web chat listing.
- [[methodology]]: weight, KV and overhead formulas.

## Related

- [[best-llm-for-128gb-mac]]
- [[mac-studio-llm]]
- [[mac-mini-llm]]
- [[how-much-ram-for-70b-model]]
- [[giant-moe-on-mac]]
- [[how-much-ram-for-local-llm]]
- [[methodology]]