Qwen2.5-Coder Code Generation (Alibaba 2024年)（クウェン2.5コーダー）
2024年Alibaba Qwen発表Qwen2.5-Coder・Industry-leading Alibaba code generation LLM + Industry-leading 0.5B-32B 6 sizes + Industry-leading 92+ programming languages + Industry-leading HumanEval 92.7% 32B-Instruct。
概要
Qwen2.5-Coder Code Generation は、2024年Alibaba Qwen発表Qwen2.5-Coder・Industry-leading Alibaba code generation LLM 2024年 + Industry-leading HumanEval 92.7% 32B-Instruct position確立。Qwen2.5-Coder specifications = Industry-leading Alibaba code generation LLM (Industry-leading Qwen2.5-Coder 2024 + Industry-leading Alibaba code generation + Industry-leading Qwen2.5-Coder flagship) + Industry-leading 0.5B-32B 6 sizes (Industry-leading 0.5B + 1.5B + 3B + 7B + 14B + 32B 6 sizes) + Industry-leading 92+ programming languages + Industry-leading HumanEval 92.7% 32B-Instruct (Industry-leading 92.7% 32B-Instruct highest)。
主な特徴・仕組み
- Org: Industry-leading Alibaba Qwen Team (Tongyi Lab)
- Year: 2024年 (arXiv 2024年9月発表 + November Qwen2.5-Coder full release)
- Paper: Industry-leading "Qwen2.5-Coder Technical Report" arXiv 2409.12186
- Sizes: Industry-leading 0.5B + 1.5B + 3B + 7B + 14B + 32B 6-tier sizes
- Languages: Industry-leading 92+ programming languages support
- Training Tokens: Industry-leading 5.5T tokens training
- Context Length: Industry-leading 128K tokens context
- HumanEval: Industry-leading HumanEval pass@1 92.7% Qwen2.5-Coder-32B-Instruct (highest in 5-tier 2024)
- MBPP: Industry-leading MBPP pass@1 90.2% 32B-Instruct
- License: Industry-leading Apache 2.0 license (commercial use OK)
- Variants: Industry-leading base + Instruct variants per size
- HuggingFace: Industry-leading Qwen/Qwen2.5-Coder-32B-Instruct
- Industry-Leading: Industry-leading Alibaba Qwen2.5-Coder 32B-Instruct highest 2024
スペック比較表
| LLM Code Generation (2023-2024) | Year | Org | Sizes | Industry Position |
|---|---|---|---|---|
| Qwen2.5-Coder | 2024 | Alibaba Qwen | 0.5B/1.5B/3B/7B/14B/32B | Industry-leading Qwen2.5-Coder 92.7% |
| DeepSeek-Coder | 2023 | DeepSeek AI | 1.3B-33B + V2 236B MoE | Industry-leading DeepSeek-Coder V2 |
| StarCoder 2 | 2024 | BigCode HF+ServiceNow | 3B/7B/15B 619 langs | Industry-leading BigCode open |
| CodeLlama | 2023 | Meta AI | 7B/13B/34B/70B | Industry-leading Meta CodeLlama |
| Codestral | 2024 | Mistral AI | 22B / Mamba 7B | Industry-leading Mistral Codestral |
具体例・対応製品
- Qwen2.5-Coder (2024年, Alibaba Qwen): Industry-leading Alibaba code generation LLM
- Industry-leading 0.5B-32B 6-tier sizes: Industry-leading 0.5B-32B 6 sizes
- Industry-leading 92+ programming languages: Industry-leading 92+ langs
- Industry-leading HumanEval 92.7% 32B-Instruct highest 2024: Industry-leading HumanEval 92.7% highest
- Industry-leading Apache 2.0 commercial license: Industry-leading Apache 2.0
- 競合 StarCoder 2 + CodeLlama + DeepSeek-Coder + Codestral: Industry-leading code gen competitors
自作PCでの選び方・注意点
Qwen2.5-Coder Code Generation は「Industry-leading Alibaba code generation LLM + Industry-leading 0.5B-32B 6 sizes」「Industry-leading 92+ programming languages + Industry-leading HumanEval 92.7% 32B-Instruct」用途のIndustry-leading Alibaba code generation 2024年Alibaba Qwen発表product。Industry-leading 0.5B-32B 6 sizes (Industry-leading 0.5B + 1.5B + 3B + 7B + 14B + 32B 6-tier sizes + Industry-leading Qwen2.5-Coder 6-tier signature + Industry-leading widest size range 5-tier) で Industry-leading 6 sizes + Industry-leading Qwen2.5-Coder signature + Industry-leading widest size range。Industry-leading HumanEval 92.7% (Industry-leading HumanEval pass@1 92.7% Qwen2.5-Coder-32B-Instruct highest in 5-tier 2024 + Industry-leading 32B-Instruct performance + Industry-leading MBPP 90.2% 32B-Instruct) で Industry-leading HumanEval 92.7% highest + Industry-leading 32B-Instruct + Industry-leading MBPP 90.2%。Industry-leading 128K context (Industry-leading 128K tokens context + Industry-leading Qwen2.5-Coder long context signature + Industry-leading 5.5T tokens training) で Industry-leading 128K + Industry-leading long context signature + Industry-leading 5.5T tokens。Industry-leading Apache 2.0 (Industry-leading Apache 2.0 license commercial use OK + Industry-leading Qwen2.5-Coder Apache signature + Industry-leading most permissive license) で Industry-leading Apache 2.0 + Industry-leading Qwen2.5 Apache + Industry-leading most permissive。Industry-leading Alibaba Qwen Tongyi (Industry-leading Alibaba Qwen Team Tongyi Lab + Industry-leading Qwen/Qwen2.5-Coder-32B-Instruct HuggingFace + Industry-leading Alibaba Qwen2.5-Coder 32B-Instruct highest 2024) で Industry-leading Alibaba Qwen Tongyi Lab + Industry-leading Qwen/Qwen2.5-Coder HuggingFace + Industry-leading 32B-Instruct highest 2024。但しIndustry-leading StarCoder 2 + CodeLlama + DeepSeek-Coder + Codestral competition (Industry-leading StarCoder 2 BigCode 619 langs 2024 + CodeLlama Meta 7B-70B 2023 + DeepSeek-Coder V2 MoE 236B 2023 + Codestral Mistral 22B 2024 vs Qwen2.5-Coder Alibaba 0.5B-32B 92.7% HumanEval Apache 2.0 2024 trade-off) で Industry-leading 4-code generation model competitors vs Qwen2.5-Coder + Industry-leading 0.5B-32B 6-tier widest range + 92+ programming languages + 5.5T tokens + 128K context + HumanEval 92.7% highest 2024 + Apache 2.0 + Alibaba Qwen Tongyi Lab 2024 unique advantage adoption alignment必須。
関連用語との違い
- vs StarCoder 2 (BigCode 2024): Qwen2.5-Coderは0.5B-32B 6 sizes + 92.7% Apache 2.0・StarCoder 2は3B-15B + 619 langs + Open RAIL-M
- vs DeepSeek-Coder V2 (236B MoE 2024): Qwen2.5-Coderは0.5B-32B dense 6 sizes・DeepSeek-Coder V2はMoE 236B + 338 langs
- vs CodeLlama/Codestral (2023-2024): Qwen2.5-Coderは6 sizes + Apache 2.0 + 92.7%・CodeLlama/Codestralはdense + Meta/Mistral
よくある質問（FAQ）
Q1: Qwen2.5-Coder vs DeepSeek-Coder 違いは？ A: Qwen2.5-Coder (Industry-leading 0.5B + 1.5B + 3B + 7B + 14B + 32B 6-tier dense + 92+ programming languages + 5.5T tokens + 128K context + HumanEval pass@1 92.7% 32B-Instruct highest + MBPP 90.2% + Apache 2.0 + Alibaba Qwen Tongyi Lab 2024) vs DeepSeek-Coder (Industry-leading 1.3B-33B V1 + V2 236B MoE 21B active + V2-Lite 16B + 86 langs V1 / 338 V2 + 6T tokens project-level + 16K-128K + HumanEval 90.2% V2 + MIT + DeepSeek AI 2023)・Industry-leading 0.5B-32B 6 sizes + 92.7% + Apache 2.0 + Alibaba = Qwen2.5-Coder + Industry-leading V2 MoE 236B + 90.2% + MIT + DeepSeek = DeepSeek-Coder preference judgment。
Q2: Industry-leading 0.5B-32B 6 sizes + 92.7% HumanEval value は？ A: Industry-leading 6 sizes + 92.7% (Industry-leading 0.5B + 1.5B + 3B + 7B + 14B + 32B 6-tier sizes + Industry-leading widest size range 5-tier + Industry-leading HumanEval pass@1 92.7% 32B-Instruct highest in 5-tier 2024 + Industry-leading MBPP 90.2%)。
Q3: Industry-leading 128K + Apache 2.0 + Tongyi Lab value は？ A: Industry-leading 128K + Apache + Tongyi (Industry-leading 128K tokens context + Industry-leading Apache 2.0 license commercial use OK + Industry-leading most permissive license + Industry-leading Alibaba Qwen Team Tongyi Lab + Industry-leading 5.5T tokens training)。
まとめ
Qwen2.5-Coder = 2024年Alibaba Qwen Tongyi Lab発表のAlibaba code generation LLM。Industry-leading 0.5B + 1.5B + 3B + 7B + 14B + 32B 6-tier dense sizes + Industry-leading widest size range 5-tier + Industry-leading 92+ programming languages support + Industry-leading 5.5T tokens training + Industry-leading 128K tokens context + Industry-leading HumanEval pass@1 92.7% Qwen2.5-Coder-32B-Instruct highest in 5-tier 2024 + Industry-leading MBPP pass@1 90.2% 32B-Instruct + Industry-leading Apache 2.0 license commercial use OK + Industry-leading base + Instruct variants per size + Industry-leading Alibaba Qwen Team Tongyi Lab + Industry-leading Qwen/Qwen2.5-Coder-32B-Instruct HuggingFace + Industry-leading Alibaba Qwen2.5-Coder 32B-Instruct highest 2024 position確立。