FitMyLLM
DeepSeek/Mixture of Experts

DeepSeekDeepSeek R1-0528

The May 2025 R1 refresh. AIME 2025 jumps from 70 to 87.5 and hallucinations drop sharply vs the January release.

chatreasoningcodingmathmultilingual
684.5B
Parameters (37B active)
160K
Context length
13
Benchmarks
17
Quantizations
713K
HF downloads
Architecture
MoE
Released
2025-05-28
Layers
61
KV Heads
128
Head Dim
56
Family
deepseek

Quantization Options

Context length:
QuantBitsVRAM @ 16KQuality
IQ2_XXS2.38
204.9 GB
204.1 + 0.8 KV
low
IQ2_M2.93
252.0 GB
251.2 + 0.8 KV
low
Q2_K3.16
271.7 GB
270.9 + 0.8 KV
low
IQ3_XXS3.25
279.4 GB
278.6 + 0.8 KV
low
IQ3_XS3.5
300.8 GB
300.0 + 0.8 KV
low
Q3_K_S3.64
312.7 GB
311.9 + 0.8 KV
low
IQ3_M3.76
323.0 GB
322.2 + 0.8 KV
low
Q3_K_M4
343.5 GB
342.7 + 0.8 KV
low
Q3_K_L4.3
369.2 GB
368.4 + 0.8 KV
moderate
IQ4_XS4.46
382.9 GB
382.1 + 0.8 KV
moderate
Q4_K_S4.67
400.9 GB
400.1 + 0.8 KV
moderate
Q4_K_M4.89
419.7 GB
418.9 + 0.8 KV
good
Q5_K_S5.57
477.9 GB
477.1 + 0.8 KV
good
Q5_K_M5.7
489.0 GB
488.2 + 0.8 KV
good
Q6_K6.56
562.6 GB
561.8 + 0.8 KV
excellent
Q8_08.5
728.6 GB
727.8 + 0.8 KV
lossless
FP1616
1370.3 GB
1369.5 + 0.8 KV
lossless

Select your GPU above to see speed estimates and compatibility for each quantization.

Too big for a single GPU — plan a multi-GPU deployment
Even the lightest quant needs ~205 GB. Size GPUs, replicas, TCO and scaling for a production setup. Open in Enterprise →
▸ READY TO RUN THIS?RENT BY THE HOUR

RENT A GPU AND RUN DEEPSEEK R1-0528 NOW

Rent on Vast.ai →Or RunPod →

Spin up an A100 / H100 / 4090 in ~60s. Pay by the second. Cancel anytime.

Community Ratings

Loading ratings...

Benchmarks (13)

AIME87.5
MMLU-PRO85.0
GPQA Diamond81.0
AA Math76.0
LiveCodeBench73.3
SWE-bench57.6
AA Long Context54.7
SciCode40.3
IFBench39.6
τ²-Bench36.5
AA Intelligence20.1
HLE17.7
Terminal-Bench15.9

Run this model

▸Easiest way to get started·Beginners
DOCS ↗
curl -fsSL https://ollama.com/install.sh | sh
$ollama run deepseek-r1:671b-0528-q4_K_M

Downloads and runs automatically. Add --verbose for speed stats.

▸ SETUP GUIDE
>_

Auto-setup with fitmyllm CLI

Detects your GPU, recommends the best model, downloads it, and starts chatting — zero config. Benchmarks your speed and contributes anonymous data to improve predictions.

pip install fitmyllmthen run fitmyllmLearn more
Auto-detect GPULive tok/s in chatSpeed benchmarks9 inference engines
HuggingFaceOllama LibraryGGUF DownloadsBuild Hardware

Find the best GPU for DeepSeek R1-0528

Build Hardware for DeepSeek R1-0528
▸ SPEC SHEET

DeepSeek R1-0528 — 684.5B MoE.

▸ SPECIFICATIONS
PARAMETERS
684.5B (37B active)
ARCHITECTURE
Mixture of Experts
CONTEXT LENGTH
160K tokens
CAPABILITIES
chat, reasoning, coding, math, multilingual
RELEASE DATE
2025-05-28
PROVIDER
DeepSeek
FAMILY
deepseek
▸ VRAM REQUIREMENTS
QUANTBPWVRAMQUALITY
IQ2_XXS2.38204.1 GB65%
IQ2_M2.93251.2 GB75%
Q2_K3.16270.9 GB78%
IQ3_XXS3.25278.6 GB82%
IQ3_XS3.5300.0 GB84%
Q3_K_S3.64311.9 GB85%
IQ3_M3.76322.2 GB86%
Q3_K_M4342.7 GB88%
Q3_K_L4.3368.4 GB90%
IQ4_XS4.46382.1 GB92%
Q4_K_S4.67400.1 GB93%
Q4_K_M4.89418.9 GB94%
Q5_K_S5.57477.1 GB96%
Q5_K_M5.7488.2 GB96%
Q6_K6.56561.8 GB97%
Q8_08.5727.8 GB100%
FP16161369.5 GB100%
§ 01BENCHMARK SCORES
MMLU-PRO85.0
aa_ifbench39.6
aa_terminal_bench15.9
aa_tau236.5
aa_scicode40.3
aa_lcr54.7
LiveCodeBench73.3
SWE-bench57.6
AIME87.5
GPQA Diamond81.0
HLE17.7
AA Intelligence20.1
AA Math76.0
§ 02RUN COMMAND

Run DeepSeek R1-0528 locally with Ollama — needs 418.9 GB VRAM at Q4_K_M:

$ollama run deepseek-r1:671b-0528