FitMyLLM
DeepSeek/Mixture of Experts

DeepSeekDeepSeek V3.2-Speciale

Reasoning-optimized variant of V3.2. Extended thinking for complex math and code tasks.

chatcodingreasoningtool_use
671B
Parameters (37B active)
128K
Context length
16
Benchmarks
17
Quantizations
Architecture
MoE
Released
2025-12-01
Layers
61
KV Heads
128
Head Dim
128
Family
deepseek

Quantization Options

Context length:
QuantBitsVRAM @ 16KQuality
IQ2_XXS2.38
200.9 GB
200.1 + 0.8 KV
low
IQ2_M2.93
247.0 GB
246.2 + 0.8 KV
low
Q2_K3.16
266.3 GB
265.5 + 0.8 KV
low
IQ3_XXS3.25
273.9 GB
273.1 + 0.8 KV
low
IQ3_XS3.5
294.9 GB
294.1 + 0.8 KV
low
Q3_K_S3.64
306.6 GB
305.8 + 0.8 KV
low
IQ3_M3.76
316.7 GB
315.9 + 0.8 KV
low
Q3_K_M4
336.8 GB
336.0 + 0.8 KV
low
Q3_K_L4.3
362.0 GB
361.2 + 0.8 KV
moderate
IQ4_XS4.46
375.4 GB
374.6 + 0.8 KV
moderate
Q4_K_S4.67
393.0 GB
392.2 + 0.8 KV
moderate
Q4_K_M4.89
411.4 GB
410.6 + 0.8 KV
good
Q5_K_S5.57
468.5 GB
467.7 + 0.8 KV
good
Q5_K_M5.7
479.4 GB
478.6 + 0.8 KV
good
Q6_K6.56
551.5 GB
550.7 + 0.8 KV
excellent
Q8_08.5
714.2 GB
713.4 + 0.8 KV
lossless
FP1616
1343.3 GB
1342.5 + 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 ~201 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 V3.2-Speciale now

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

Community Ratings

Loading ratings...

Benchmarks (16)

Run this model

Easiest way to get started·Beginners
Docs ↗
curl -fsSL https://ollama.com/install.sh | sh
$# DeepSeek V3.2-Speciale is not in Ollama's library, and we know no GGUF build of it

Search Hugging Face for a GGUF of it (https://huggingface.co/models?sort=trending&search=DeepSeek-V3.2-Speciale+GGUF), or use the vLLM tab with the original weights.

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

Find the best GPU for DeepSeek V3.2-Speciale

Build Hardware for DeepSeek V3.2-Speciale
Spec sheet

DeepSeek V3.2-Speciale — 671B MoE.

Specifications
Parameters
671B (37B active)
Architecture
Mixture of Experts
Context Length
128K tokens
Capabilities
chat, coding, reasoning, tool_use
Release Date
2025-12-01
Provider
DeepSeek
Family
deepseek
VRAM requirements
QuantBPWVRAMQuality
IQ2_XXS2.38200.1 GB65%
IQ2_M2.93246.2 GB75%
Q2_K3.16265.5 GB78%
IQ3_XXS3.25273.1 GB82%
IQ3_XS3.5294.1 GB84%
Q3_K_S3.64305.8 GB85%
IQ3_M3.76315.9 GB86%
Q3_K_M4336.0 GB88%
Q3_K_L4.3361.2 GB90%
IQ4_XS4.46374.6 GB92%
Q4_K_S4.67392.2 GB93%
Q4_K_M4.89410.6 GB94%
Q5_K_S5.57467.7 GB96%
Q5_K_M5.7478.6 GB96%
Q6_K6.56550.7 GB97%
Q8_08.5713.4 GB100%
FP16161342.5 GB100%
Benchmark scores
MMLU-PRO85.0
BigCodeBench50.0
LiveCodeBench89.6
SWE-bench70.0
AIME97.0
MATH-50096.7
GPQA Diamond82.4
HLE26.1
AA Intelligence29.4
AA Coding37.9
AA Math96.7
IFBench63.9
Terminal-Bench Hard34.8
SciCode44.0
AA Long Context59.3
EQ-Bench Creative Writing v31277