Model

Qwen2.5-32B

OPEN

Alibaba · Qwen2.5 family · released Sep 19, 2024

Popular Apache-2.0 quality-per-cost dense model; officially reported to beat larger Gemma2-27B/Phi-3.5-MoE baselines.

ReasoningCodingVisionFunction callingTool useAgentic

843.7

Elo · rank #246

Parameters

32.5B

Active params

32.5B (dense)

Context

128K tokens

Architecture

Dense transformer, 64 layers (GQA 40Q/8KV)

License

Apache-2.0

Languages

29+

API price (in/out)

No hosted API

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

ARC-ChallengeReasoning70.4%#48

best: Llama 3.1 405B · 96.9%

BIG-Bench HardReasoning84.5%#19

best: ERNIE 4.5 300B-A47B · 94.3%

GPQA DiamondReasoning48.0%#194

best: GPT-5.6 · 94.6%

GSM8KMath92.9%#24

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning85.2%#40

best: Claude 3 Opus · 95.4%

HumanEval+Coding52.4%#26

best: Mistral Small 3.2 (24B) · 92.9%

HumanEvalCoding58.5%#106

best: Claude Opus 4.5 · 99.4%

MATH-500Math57.7%#116

best: GPT-5 · 99.4%

MBPP+Coding67.2%#21

best: Llama 3.1 405B · 88.6%

MBPPCoding84.5%#6

best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%

MMLU-ProKnowledge55.1%#123

best: Claude Fable 5 · 91.5%

MMLU-ReduxKnowledge82.0%#33

best: Qwen3.7-Max · 95.0%

MMLU (STEM)Knowledge80.9%#3

best: Falcon-H1 34B · 83.6%

MMLUKnowledge83.3%#71

best: OpenAI o3 · 92.9%

TruthfulQAKnowledge57.8%#30

best: Phi-3.5-MoE (16x3.8B, 6.6B active) · 77.5%

WinoGrandeReasoning82.0%#23

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

19.9 GB

VRAM @ FP16

65 GB

Fits on (Q4)

RTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

~20GB Q4 fits within 24GB with limited headroom for context/KV cache; no single independently verified 4090 tok/s figure found.

Quantizations

GGUF · AWQ · GPTQ Int4 · MLX

Qwen2.5 family

Elo progression across releases

Qwen2.5-14BSep 2024696.7 Qwen2.5-32BSep 2024843.7 Qwen2.5-72BSep 2024965.7 Qwen2.5-7BSep 2024565.7 Qwen2.5-MaxJan 20251252.5

Resources

HFQwen/Qwen2.5-32B GHQwenLM/Qwen2.5 DOCSqwenlm.github.io/blog/qwen2.5

API price weights · each benchmark row carries its own source badge (see methodology)