Model

Qwen2-Math-72B

OPEN

Alibaba · Qwen2-Math family · released Aug 8, 2024

First Qwen math specialist; Instruct variant beat GPT-4o and Claude-3.5-Sonnet on math benchmarks.

ReasoningCodingVisionFunction callingTool useAgentic

1145.4

Elo · rank #201

Parameters

72.7B

Active params

72.7B (dense)

Context

4K tokens

Architecture

Dense decoder-only Transformer (GQA), math-specialized fine-tune of Qwen2-72B

License

Tongyi Qianwen License Agreement

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

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

AIMEMath20.0%#142

best: GPT-5.2 · 100.0%

AMCMath60.0%#2

best: Falcon-H1 34B · 69.4%

Gaokao MathMath59.8%#4

best: Doubao-Seed-1.6 · 96.0%

GSM8KMath96.7%#2

best: Llama 3.1 405B · 96.8%

MATH-500Math84.0%#58

best: GPT-5 · 99.4%

MMLU (STEM)Knowledge79.9%#5

best: Falcon-H1 34B · 83.6%

OlympiadBenchMath43.0%#3

best: Doubao-1.5-Pro · 59.8%

Run it locally

VRAM @ Q4

42 GB

VRAM @ FP16

145 GB

Fits on (Q4)

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

72B dense weights exceed a single RTX 4090's 24GB VRAM even at Q4.

Quantizations

GGUF · AWQ · GPTQ

Fine-tune it

Conditional / custom

QLoRA49.4 GB1× A100 80GB

LoRA154.9 GB1× B200 192GB

Full fine-tune1166.8 GB8× B200 192GB

QLoRA SFT on ~10k samples ≈ $51.13 (1× A100 80GB)

Plan a fine-tune →

Qwen2-Math family

Elo progression across releases

Qwen2-Math-72BAug 20241145.4

Resources

HFQwen/Qwen2-Math-72B-Instruct GHQwenLM/Qwen2-Math DOCSqwenlm.github.io/blog/qwen2-math

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