Model

Qwen2-57B-A14B

OPEN

Alibaba · Qwen2 family · released Jun 7, 2024

Gen-2 MoE (14B active of 57B total) matching ~30B-dense-class quality; Apache 2.0.

ReasoningCodingVisionFunction callingTool useAgentic

589.2

Elo · rank #283

Parameters

57B

Active params

14B (MoE)

Context

64K tokens

Architecture

Mixture-of-Experts Transformer decoder (57B total / 14B active per token)

License

Apache 2.0

Languages

30+

API price (in/out)

No hosted API

Modalities

text

Benchmark results

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

ARC-ChallengeReasoning64.1%#72

best: Llama 3.1 405B · 96.9%

BIG-Bench HardReasoning67.0%#67

best: ERNIE 4.5 300B-A47B · 94.3%

C-EvalKnowledge87.7%#18

best: Qwen3.6-Plus · 93.3%

GPQA DiamondReasoning34.3%#244

best: GPT-5.6 · 94.6%

GSM8KMath80.7%#81

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning85.2%#41

best: Claude 3 Opus · 95.4%

HumanEvalCoding53.0%#116

best: Claude Opus 4.5 · 99.4%

IFEvalReasoning59.9%#121

best: Gemma 4 26B A4B · 98.5%

LiveCodeBenchCoding25.5%#159

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

LogicKorHuman preference6.6#24

best: GPT-4o · 9.33

MATH-500Math43.0%#146

best: GPT-5 · 99.4%

MBPPCoding71.9%#34

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

MGSMMath74.0%#31

best: OpenAI o4-mini · 93.7%

MMLU-ProKnowledge43.0%#147

best: Claude Fable 5 · 91.5%

MMLUKnowledge76.5%#119

best: OpenAI o3 · 92.9%

MT-BenchHuman preference8.55#24

best: Hunyuan-Large (A52B) · 9.4

TruthfulQAKnowledge57.7%#31

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

WinoGrandeReasoning79.5%#40

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

32 GB

VRAM @ FP16

115 GB

Fits on (Q4)

M4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

57B total weights exceed a single RTX 4090's 24GB VRAM even at Q4.

Quantizations

GGUF · AWQ · GPTQ

Qwen2 family

Elo progression across releases

Qwen2-57B-A14BJun 2024589.2 Qwen2-72BJun 2024832.7 Qwen2-7BJun 2024485.5

Resources

HFQwen/Qwen2-57B-A14B-Instruct GHQwenLM/Qwen2 DOCSqwenlm.github.io/blog/qwen2

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