Qwen 2.5 72B

Best Coding Open Source

Alibaba Cloud

Released: 2024-09-19

Type: Open Weights LLM

Context

128k

Max Out

8.2k

Cutoff

2024-08

Params

72B

The Breakdown

Qwen 2.5 72B is the model that made the West pay attention to Alibaba's AI labs. It fundamentally shifted the open-weights leaderboard, overtaking Meta's Llama 3.1 in crucial coding and mathematics benchmarks. For developers building code-generation assistants or complex logic agents, Qwen offers a level of precision that feels 'smarter' than its parameter count suggests. While users should be aware of its alignment constraints regarding specific political topics, its raw reasoning engine is world-class.

Overall Score

9.1

/10

Pricing (per 1k tokens)

Input$0

Output$0

Currency: USD (Self-Hosted)

The Good

Incredible coding abilities (beats Llama 3.1 in many benchmarks)
Strong multilingual support, especially Asian languages
Apache 2.0 license allows broad commercial use

The Bad

Heavier censorship on politically sensitive topics (China-origin)
Requires significant VRAM to run locally (2x A100s preferred)
Brand familiarity is lower in Western enterprise

The Verdict

The sleeper hit of 2024. Qwen 2.5 consistently outperforms Llama 3 70B in coding and math. If you are self-hosting for a dev tool or code agent, this is likely the model you want, despite the geopolitical caveats.

Performance Benchmarks

human eval

89.6

drop

82.5

mmlu