Model

Qwen-72B

OPEN

Alibaba · Qwen family · released Nov 30, 2023

Gen-1 flagship trained on 3T tokens with 32K context; beat Llama2-70B and topped GPT-3.5 on MMLU/C-Eval/GSM8K at release.

ReasoningCodingVisionFunction callingTool useAgentic

535.8

Elo · rank #293

Parameters

72B

Active params

72B (dense)

Context

32K tokens

Architecture

Dense decoder-only transformer, 80 layers (RoPE, SwiGLU, RMSNorm)

License

Tongyi Qianwen LICENSE AGREEMENT (custom, free commercial use below usage threshold)

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

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

BIG-Bench HardReasoning67.7%#65

best: ERNIE 4.5 300B-A47B · 94.3%

C-EvalKnowledge83.3%#27

best: Qwen3.6-Plus · 93.3%

CMMLUKnowledge83.6%#9

best: Doubao-1.5-Pro · 90.9%

GSM8KMath78.9%#88

best: Llama 3.1 405B · 96.8%

HumanEvalCoding35.4%#142

best: Claude Opus 4.5 · 99.4%

LogicKorHuman preference6.73#22

best: GPT-4o · 9.33

MATH-500Math35.2%#152

best: GPT-5 · 99.4%

MBPPCoding52.2%#71

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

MMLUKnowledge77.4%#111

best: OpenAI o3 · 92.9%

Run it locally

VRAM @ Q4

48 GB

VRAM @ FP16

144 GB

Fits on (Q4)

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

Q4 weights (~48GB) exceed a single RTX 4090's 24GB VRAM; requires multi-GPU or CPU/disk offload.

Quantizations

GPTQ Int4 · GPTQ Int8 · GGUF

Qwen family

Elo progression across releases

Qwen-7BAug 2023137.5 Qwen-14BSep 2023347.6 Qwen-1.8BNov 2023-157.9 Qwen-72BNov 2023535.8

Resources

HFQwen/Qwen-72B GHQwenLM/Qwen

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