back to local model checker

DeepSeek R1

MoEThinkingPopular

DeepSeek · 671B (37B active) · Mixture of Experts

Massive MoE reasoning model — 37B active

MIT
Downloads

1.1M

Likes

13.1K

Released

2025-01

Context

64K tokens

Use cases

reasoning
Total experts256
Active experts8
Active params37.0B

Your hardware

16 GB Reference Desktop

Quant
Q4_K_M
Memory
344.2 GB
Speed
--
Fit
Too heavy

Quantization options

What is quantization?
QuantBitsVRAMDiskQualityStatus
Q2_K2215.3 GB205.1 GBlowToo heavy · ~1 t/s
Q3_K_M3301.2 GB287.1 GBmoderateToo heavy · ~1 t/s
Q4_K_M4344.2 GB328.1 GBgoodToo heavy · ~1 t/s
Q5_K_M5430.1 GB410.1 GBgoodToo heavy · ~1 t/s
Q6_K6516.1 GB492.1 GBexcellentToo heavy · ~1 t/s
Q8_08687.9 GB656.2 GBexcellentToo heavy · ~1 t/s
F16161375.3 GB1312.3 GBlosslessToo heavy · ~1 t/s

About this model

When sizing hardware for DeepSeek r1 model download (DeepSeek R1), deepSeek R1 is the full DeepSeek reasoning model, not one of the smaller distilled variants. It has 671B total parameters and 37B active parameters in this data set, with a heavyweight mixture-of-experts architecture. It is meant for reasoning-heavy workloads on very high-memory local or server-class hardware.

Most laptops and desktops should start with a DeepSeek R1 Distill model before attempting the full model. The 7B, 14B, and 32B distills are easier to keep resident and are more practical for local testing. Full R1 should be treated as a capacity planning problem first.

Highlights

  • Full reasoning modelThis is the full DeepSeek R1 MoE, not a smaller distillation.
  • Huge memory targetFull R1 is a server-class local run for most people, even when quantized.
  • Distills matterMost desktop users should test DeepSeek R1 Distill models before attempting full R1.
  • Reasoning baselineUse it as the heavyweight reasoning reference when comparing local Gemma, Qwen, and DeepSeek options.

System requirements by device

Memory planning targets 375 GB minimum RAM/VRAM and 624.9 GB recommended memory. The baseline quant is Q4_K_M, estimated at 344.2 GB memory and 328.1 GB disk. Actual runtime usage shifts with context length, KV cache precision, and OS background overhead.

DeviceNotes
Mac (Apple silicon)Requires a 48 GB–192 GB Mac Studio or MacBook Pro Max/Ultra with enough unified wired budget for 344.2 GB.
Windows / Linux GPUUse dual/multi-GPU CUDA splitting (--tensor-split) or MoE CPU expert offloading (-ot) paired with 32–128 GB DDR5 system RAM.
Phones & Compact Edge BoardsNot practical on phones (375 GB min RAM exceeds Android LMK and iOS Jetsam limits). Pick a 2B–4B model on mobile.

Start with a short session and leave memory for the operating system. A file fitting on disk does not guarantee that it will load in memory.

Estimated memory by quantization

These are rough estimates, not tested minimum requirements. Check the actual file and runtime documentation; context, offloading and operating-system overhead can change the memory needed.

Quant TierEstimated memoryDisk SizeQuality
Q2_K215.3 GB205.1 GBlow
Q3_K_M301.2 GB287.1 GBmoderate
Q4_K_M344.2 GB328.1 GBgood
Q5_K_M430.1 GB410.1 GBgood
Q6_K516.1 GB492.1 GBexcellent
Q8_0687.9 GB656.2 GBexcellent
F161375.3 GB1312.3 GBlossless

How to run DeepSeek R1 on your device

Start with a short prompt and the smallest quant that still gives the quality you need. A clean first run is more useful than loading a larger file with no memory headroom.

Mac

Check DeepSeek R1 on Mac with LM Studio

  1. Open LM Studio on Mac and search for DeepSeek R1 GGUF.
  2. Check the full model file size before download.
  3. Download a DeepSeek R1 Distill model first if unified memory is not workstation-class.
  4. Run a short reasoning prompt before increasing context.
  5. Stop and use DeepSeek R1 Distill 32B if the full model swaps.
Device how-to guide
Windows

Check DeepSeek R1 on Windows with Ollama or LM Studio

  1. Open LM Studio or Ollama on Windows and search for DeepSeek R1.
  2. Verify RAM, VRAM, and disk space before downloading the full model.
  3. Start with deepseek-r1:32b or another distill if the full model is too large.
  4. Run one short math or reasoning prompt.
  5. Only raise context after the first answer streams without paging.
Device how-to guide
Android

Use a DeepSeek R1 Distill on Android in AI Edge Gallery

  1. Install AI Edge Gallery and open the local model picker.
  2. Do not import the full DeepSeek R1 model on a normal phone.
  3. Choose a small DeepSeek R1 distill such as 1.5B or 7B if available.
  4. Keep context short and test one reasoning prompt.
  5. Move the workflow to desktop if Android reloads the model.
Device how-to guide
iPhone

Use a DeepSeek R1 Distill on iPhone in PocketPal

  1. Install PocketPal and open the model import screen.
  2. Skip full DeepSeek R1 on iPhone unless you are only testing feasibility.
  3. Import a small DeepSeek R1 distill through Files.
  4. Run one short reasoning prompt with low context.
  5. Use a Mac or Windows machine for larger DeepSeek R1 experiments.
Device how-to guide

DeepSeek R1 vs alternatives

ModelParamsContextMin RAMLicense
DeepSeek R1671B (37B active)64K375 GBMIT
DeepSeek R1 Distill 32B32B64K17.9 GBMIT
Qwen 3 235B-A22B235B (22B active)128K131.3 GBApache 2.0
Qwen 3 30B-A3B30B (3.3B active)128K16.8 GBApache 2.0
Gemma 4 26B-A4B IT27B (4B active)256K15.1 GBGemma

Use this table to choose the smallest nearby model that still matches the job. A related model with lower RAM can be the better local choice when the current page fits only at the edge of available memory.

Troubleshooting & performance tips

Full R1 will not fit

That is the expected failure on most local machines. Use a DeepSeek R1 distill, lower context, or move the full model to a server-class setup.

Reasoning is too slow

DeepSeek R1 spends tokens on reasoning. Shorten the prompt, reduce context, and compare against a distill before changing runtimes.

Ollama downloads the wrong size

Check the exact tag before download. A distill and the full DeepSeek R1 model have very different memory profiles even when the name looks similar.

Phone setup fails

Use a distill on phones. Full DeepSeek R1 is not a practical Android or iPhone target for normal local inference.

Frequently asked questions

Can I run DeepSeek R1 locally?
Yes, you can run DeepSeek R1 locally only on very high-memory hardware if you mean the full model. Most people should run a DeepSeek R1 distill locally instead.
What is DeepSeek R1 0528?
DeepSeek R1 0528 is a later DeepSeek R1 release. Check the exact model card and file before treating it as the same local download.
Can DeepSeek R1 run on Mac?
DeepSeek R1 can run on Mac only when the Mac has enough unified memory for the chosen full or distilled model. For normal deepseek r1 mac use, start with a distill.
Is DeepSeek R1 good for reasoning?
Yes, DeepSeek R1 is built for reasoning and is the heavyweight reference in this model set. The cost is memory, latency, and setup complexity compared with smaller local models.
Should I use full DeepSeek R1 or a distill?
Use a DeepSeek R1 distill unless you have server-class local hardware. Distills are the practical answer for laptops, desktops, and phone experiments.
DeepSeek R1 vs Qwen3: which is better locally?
Qwen3 30B A3B is usually better locally when you need a realistic desktop model. Full DeepSeek R1 is stronger as a reasoning reference but much harder to fit.
How do I download and install DeepSeek r1 model download?
Download a GGUF build for DeepSeek R1, start with Q4_K_M, and confirm the model fits locally before increasing context or quant quality.