The Open-Source AI Stack
RSS
All models

Models · deepseek

DeepSeek-R1-Distill-Qwen-32B

Open DeepSeek · 2025-01-20 · MIT

Dense Qwen2.5-32B base supervised-fine-tuned on reasoning traces sampled from DeepSeek-R1. The sweet-spot distill of the series: small enough to run on a single high-memory GPU, and reported by DeepSeek as outperforming OpenAI o1-mini on AIME 2024 and MATH-500.

Cost

$0.00 / Mtok input
$0.00 / Mtok output

Median across providers · as of 2026-05-21

via Artificial Analysis ↗

Speed

0 tok/sec output
0 ms TTFT

· as of 2026-05-21

source ↗

Architecture

tokens in Embedding vocab not disclosed · qwen2 tokenizer × N layers Grouped-Query Attention RoPE context 128,000 tokens Dense MLP SwiGLU activation (standard) 32B active params Output projection tokens out
Schema-generated from data/models.yaml. Every label is auditable against the model's sources.

Specs

Architecture
dense
Total params
32B
Active params
32B
Context window
128K tokens
Attention
gqa
Position encoding
rope
Post-training
sft
OSI-approved
yes
Data released
no
Training code
not released

Benchmarks

Each score carries the date it was published; we never infer or interpolate missing scores.

General reasoning

MMLU-Pro 73.9 as of 2026-05-21 source ↗
GPQA-Diamond 62.1 as of 2025-01-22 source ↗

Code

LiveCodeBench 57.2 as of 2025-01-22 source ↗

Math

MATH 94.1 as of 2026-05-21 source ↗
AIME 2024 72.6 as of 2025-01-22 source ↗
AIME 2025 63.0 as of 2026-05-21 source ↗

Recommended use cases

  • single-GPU open reasoning
  • math and code reasoning at local-deployable scale
  • reasoning inside Qwen-compatible stacks

Available quantizations

GGUF llama.cpp's container; the common local format, k-quants from Q2 to Q8. runs on llama.cpp, Ollama
AWQ Activation-aware 4-bit weight quantization for GPU serving. runs on vLLM, SGLang
GPTQ Post-training 4-bit weight quantization for GPU serving. runs on vLLM, SGLang, Transformers
EXL2 ExLlamaV2's variable-bitrate format for consumer GPUs. runs on ExLlamaV2
MLX Apple MLX 4/8-bit layout for Apple silicon. runs on Apple MLX
FP8 8-bit float, frequently a native release on Hopper / Blackwell GPUs. runs on vLLM, SGLang, TensorRT-LLM
bitsandbytes On-the-fly NF4 / INT8 weight quantization inside Transformers. runs on Transformers

Verified via the Hugging Face model tree ↗. Community quantizations change over time; the families shown are those with published weights at audit time.

Notable innovations

  • · MIT-licensed reasoning at 32B dense
  • · Single-GPU-deployable reasoning model
  • · Outperformed o1-mini on AIME 2024 and MATH-500 per DeepSeek's report

Lineage

Distilled from R1 into a Qwen 2.5 32B base.

Derived from

DeepSeek-R1 2025-01-20

Sources