Model

Qwen2.5-Max

CLOSED

Alibaba · Qwen2.5 family · released Jan 28, 2025

Large proprietary MoE pretrained on 20T+ tokens; API model id qwen-max-2025-01-25, competitive with DeepSeek V3.

ReasoningCodingVisionFunction callingTool useAgentic

1252.5

Elo · rank #167

Parameters

Undisclosed

Active params

Undisclosed

Context

32K tokens

Architecture

Proprietary large-scale MoE (configuration undisclosed)

License

Proprietary

Languages

—

API price (in/out)

$1.6 / $6.4

Modalities

text

Benchmark results

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

Arena-HardHuman preference89.4%#13

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

Arena EloHuman preference1374#27

best: Claude Fable 5 · 1505

C-EvalKnowledge92.2%#4

best: Qwen3.6-Plus · 93.3%

GPQA DiamondReasoning60.1%#164

best: GPT-5.6 · 94.6%

GSM8KMath94.5%#14

best: Llama 3.1 405B · 96.8%

LiveCodeBenchCoding38.7%#134

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

MATH-500Math68.5%#99

best: GPT-5 · 99.4%

MMLU-ProKnowledge76.1%#70

best: Claude Fable 5 · 91.5%

MMLUKnowledge87.9%#27

best: OpenAI o3 · 92.9%

Run it locally

Closed weights — available via API only. No local deployment.

Input / M tok

$1.6

Output / M tok

$6.4

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

DOCSqwenlm.github.io/blog/qwen2.5-max APIAvailable via Alibaba API

API price $1.6/$6.4 · each benchmark row carries its own source badge (see methodology)