Model

Qwen2.5-72B

OPEN

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

Gen-2.5 dense flagship trained on 18T tokens; large MMLU/MATH/coding gains over Qwen2; Qwen license (non-Apache) for the 72B/3B sizes.

ReasoningCodingVisionFunction callingTool useAgentic

965.7

Elo · rank #221

Parameters

72.7B

Active params

72.7B (dense)

Context

128K tokens

Architecture

Dense transformer, 80 layers (GQA 64Q/8KV)

License

Qwen (Tongyi Qianwen License)

Languages

29+

API price (in/out)

No hosted API

Modalities

text

Benchmark results

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

ARC-ChallengeReasoning72.4%#43

best: Llama 3.1 405B · 96.9%

Arena-HardHuman preference81.2%#26

best: Qwen3-235B-A22B (Non-Thinking) · 96.1%

BIG-Bench HardReasoning86.3%#14

best: ERNIE 4.5 300B-A47B · 94.3%

GPQA DiamondReasoning45.9%#204

best: GPT-5.6 · 94.6%

GSM8KMath91.5%#35

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning87.6%#16

best: Claude 3 Opus · 95.4%

HumanEval+Coding51.2%#28

best: Mistral Small 3.2 (24B) · 92.9%

HumanEvalCoding59.1%#102

best: Claude Opus 4.5 · 99.4%

IFEvalReasoning84.1%#70

best: Gemma 4 26B A4B · 98.5%

LiveCodeBenchCoding55.5%#110

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MATH-500Math62.1%#108

best: GPT-5 · 99.4%

MBPP+Coding69.2%#16

best: Llama 3.1 405B · 88.6%

MBPPCoding84.7%#5

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

MMLU-ProKnowledge58.1%#117

best: Claude Fable 5 · 91.5%

MMLU-ReduxKnowledge83.9%#32

best: Qwen3.7-Max · 95.0%

MMLUKnowledge86.1%#45

best: OpenAI o3 · 92.9%

MT-BenchHuman preference9.35#2

best: Hunyuan-Large (A52B) · 9.4

TruthfulQAKnowledge60.4%#22

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

WinoGrandeReasoning83.9%#13

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

47.4 GB

VRAM @ FP16

145 GB

Fits on (Q4)

M3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

Q4 weights (~47GB) exceed a single RTX 4090's 24GB VRAM; requires multi-GPU or CPU offload for local inference.

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/Qwen2.5-72B GHQwenLM/Qwen2.5 DOCSqwenlm.github.io/blog/qwen2.5

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