DeepSeek-R1-Distill-Qwen-32B

Name: DeepSeek-R1-Distill-Qwen-32B
Author: DeepSeek

DeepSeek · DeepSeek-R1 distill · released 2025-01-20 · data checked 2026-11-01 Needs verification

DeepSeek-R1-Distill-Qwen-32B is a 32.8B-parameter dense model from DeepSeek (China), released 2025-01-20 under the MIT license. It scores 50.9 on the Small LM Index, #28 of 34 medium models. It needs about 23.2 GB of memory at 4-bit with an 8K-token context.

Small LM Index

50.9

#28 of 34 scored medium models

Likely 46.1–55.6 · from 3 of 6 core benchmarks

Parameters

32.8B

Dense

Price per 1M tokens

—

No hosted API on OpenRouter; self-host

Context window

128K

tokens

Benchmarks

Each bar shows this model's score. The tick marks the best medium model on that benchmark. Sources: 3 reported by DeepSeek. How we label sources.

GPQA Diamond Lab-reported Knowledge62.1

Class best: Nanbeige4.3 33B 89.5

AIME Lab-reported Math · AIME 202472.6

Class best: Ternary Bonsai 2 27B 95.8

LiveCodeBench Lab-reported Coding57.2

Class best: Qwen3.8 27B 90.3

Not reported: estimated

What the index model expects from this model's other scores. Not a measurement.

MMLU-Pro≈ 73

Humanity's Last Exam≈ 7

SWE-bench Verified≈ 32

Specifications

Lab: DeepSeek (CN · China)
Architecture: Dense transformer
Total parameters: 32.8B
Active parameters: All (dense)
Size class: Medium
Context window: 128K tokens
Input modalities: text
Reasoning: Reasoning (thinking)
License: MIT
Openness: Open weights
Memory to run: ~23.2 GB at 4-bit, 8K context (more sizes)
Hugging Face: 348.2K downloads (30 days) · 1.7K likes

Weights on Hugging Face ↗

Compare with

DeepSeek-R1-Distill-Qwen-32B vs Nanbeige4.3 33B (33.1B, index 77.8)
DeepSeek-R1-Distill-Qwen-32B vs Sarvam 30B (32.2B, index 61.6)
DeepSeek-R1-Distill-Qwen-32B vs OLMo 3.1 32B Think (32.2B, index 57.0)
DeepSeek-R1-Distill-Qwen-32B vs Qwen3 32B (32.8B, index 52.4)

Sources

https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B

Something wrong here? Report it.

Run it locally

Uses unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF (Q4_K_M). Swap the tag for another size: F16, Q2_K, Q2_K_L, Q3_K_M, Q4_K_M, Q5_K_M, Q6_K, Q8_0.

Ollama

ollama run hf.co/unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M

llama.cpp

llama-server -hf unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M

Then open http://localhost:8080 for a chat window.

MLX (Mac)

pip install -U mlx-lm
mlx_lm.chat --model mlx-community/DeepSeek-R1-Distill-Qwen-32B-4bit

vLLM (GPU server)

vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-32B

Full-precision weights, for a server GPU.

Open in LM Studio

Memory needed: weights, context cache and runtime buffers.

Quantization	4K context	32K context	128K context
16-bit F16	68 GB	75 GB	100 GB
8-bit Q8_0	37 GB	45 GB	70 GB
4-bit Q4_K_M	22 GB	30 GB	55 GB

Check it against your GPU, Mac or phone →