Model

Qwen2.5-14B

OPEN

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

Apache-2.0 dense mid-size model from the 18T-token Qwen2.5 pretrain; 128K context via YaRN.

ReasoningCodingVisionFunction callingTool useAgentic

696.7

Elo · rank #267

Parameters

14.7B

Active params

14.7B (dense)

Context

128K tokens

Architecture

Dense transformer (GQA)

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-ChallengeReasoning67.3%#59

best: Llama 3.1 405B · 96.9%

BIG-Bench HardReasoning78.2%#37

best: ERNIE 4.5 300B-A47B · 94.3%

GPQA DiamondReasoning32.8%#252

best: GPT-5.6 · 94.6%

GSM8KMath90.2%#40

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning84.3%#52

best: Claude 3 Opus · 95.4%

HumanEval+Coding51.2%#27

best: Mistral Small 3.2 (24B) · 92.9%

HumanEvalCoding56.7%#110

best: Claude Opus 4.5 · 99.4%

MATH-500Math55.6%#118

best: GPT-5 · 99.4%

MBPP+Coding63.2%#26

best: Llama 3.1 405B · 88.6%

MBPPCoding76.7%#23

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

MMLU-ProKnowledge51.2%#131

best: Claude Fable 5 · 91.5%

MMLU-ReduxKnowledge76.6%#36

best: Qwen3.7-Max · 95.0%

MMLU (STEM)Knowledge76.4%#6

best: Falcon-H1 34B · 83.6%

MMLUKnowledge79.7%#92

best: OpenAI o3 · 92.9%

TruthfulQAKnowledge58.4%#27

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

WinoGrandeReasoning81.0%#33

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

9 GB

VRAM @ FP16

29 GB

Fits on (Q4)

RTX 3060 12GBRTX 4070 Ti 16GBRTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

Fits comfortably in 24GB VRAM at Q4 (~9GB); no independently verified 4090 llama.cpp 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/Qwen2.5-14B GHQwenLM/Qwen2.5 DOCSqwenlm.github.io/blog/qwen2.5

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