DeepSeek-Coder Code Generation (DeepSeek 2023年)（ディープシークコーダー）
2023年DeepSeek AI発表DeepSeek-Coder・Industry-leading China code generation LLM + Industry-leading 1.3B-33B + V2 236B MoE + Industry-leading 86 programming languages + Industry-leading project-level code dataset。
概要
DeepSeek-Coder Code Generation は、2023年DeepSeek AI発表DeepSeek-Coder・Industry-leading China code generation LLM 2023年 + Industry-leading project-level code dataset position確立。DeepSeek-Coder specifications = Industry-leading China code generation LLM (Industry-leading DeepSeek-Coder 2023 + Industry-leading China code generation + Industry-leading DeepSeek-Coder flagship) + Industry-leading 1.3B-33B (Industry-leading 1.3B + 6.7B + 33B variants) + Industry-leading V2 236B MoE (Industry-leading DeepSeek-Coder V2 236B MoE 2024) + Industry-leading 86 programming languages (Industry-leading 86 languages) + Industry-leading project-level code dataset (Industry-leading project-level code training)。
主な特徴・仕組み
- Org: Industry-leading DeepSeek AI (China・Founder Liang Wenfeng)
- Year: 2023年 (DeepSeek-Coder arXiv 2023年11月 + V2 2024年6月)
- Paper: Industry-leading "DeepSeek-Coder: When the Large Language Model Meets Programming" arXiv 2401.14196 + "DeepSeek-Coder-V2" arXiv 2406.11931
- V1 Sizes: Industry-leading 1.3B + 5.7B + 6.7B + 33B variants
- V2 Sizes: Industry-leading DeepSeek-Coder V2 236B MoE (21B active) + V2-Lite 16B (2.4B active)
- Languages: Industry-leading 86 programming languages support (V1) / 338 langs (V2)
- Dataset: Industry-leading 2T tokens (V1) / 6T tokens (V2) project-level code training
- Context Length: Industry-leading 16K-128K context
- HumanEval: Industry-leading HumanEval pass@1 90.2% DeepSeek-Coder-V2-Instruct (236B)
- MIT License: Industry-leading MIT-style commercial use license
- DeepSeek AI: Industry-leading DeepSeek AI Liang Wenfeng founded 2023
- Industry-Leading: Industry-leading DeepSeek-Coder V2 236B MoE China code generation
スペック比較表
| LLM Code Generation (2023-2024) | Year | Org | Sizes | Industry Position |
|---|---|---|---|---|
| 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 |
| Qwen2.5-Coder | 2024 | Alibaba Qwen | 0.5B/1.5B/3B/7B/14B/32B | Industry-leading Qwen2.5-Coder |
| Codestral | 2024 | Mistral AI | 22B / Mamba 7B | Industry-leading Mistral Codestral |
具体例・対応製品
- DeepSeek-Coder (2023年, DeepSeek AI): Industry-leading China code generation LLM
- Industry-leading V2 236B MoE 21B active: Industry-leading V2 236B MoE
- Industry-leading 86 programming languages V1 / 338 langs V2: Industry-leading 86 / 338 langs
- Industry-leading HumanEval pass@1 90.2% V2-Instruct 236B: Industry-leading HumanEval 90.2%
- Industry-leading 6T tokens project-level code training: Industry-leading 6T tokens project-level
- 競合 StarCoder 2 + CodeLlama + Qwen2.5-Coder + Codestral: Industry-leading code gen competitors
自作PCでの選び方・注意点
DeepSeek-Coder Code Generation は「Industry-leading China code generation LLM + Industry-leading 1.3B-33B + V2 236B MoE」「Industry-leading 86 programming languages + Industry-leading project-level code dataset」用途のIndustry-leading China code generation 2023年DeepSeek AI発表product。Industry-leading V2 236B MoE (Industry-leading DeepSeek-Coder V2 236B MoE 21B active + Industry-leading V2-Lite 16B 2.4B active + Industry-leading MoE architecture signature) で Industry-leading V2 236B MoE + Industry-leading V2-Lite 16B + Industry-leading MoE signature。Industry-leading 86/338 langs (Industry-leading 86 programming languages V1 + 338 languages V2 + Industry-leading DeepSeek-Coder multilang signature) で Industry-leading 86 V1 + 338 V2 + Industry-leading multilang signature。Industry-leading project-level (Industry-leading project-level code dataset + Industry-leading 6T tokens V2 + Industry-leading DeepSeek-Coder project-level signature) で Industry-leading project-level + Industry-leading 6T tokens V2 + Industry-leading project-level signature。Industry-leading HumanEval 90.2% (Industry-leading HumanEval pass@1 90.2% DeepSeek-Coder-V2-Instruct 236B + Industry-leading V2-Instruct performance + Industry-leading 236B flagship) で Industry-leading HumanEval 90.2% + Industry-leading V2-Instruct + Industry-leading 236B flagship。Industry-leading 16K-128K (Industry-leading 16K-128K context length + Industry-leading DeepSeek-Coder long context signature + Industry-leading V2 long context) で Industry-leading 16K-128K + Industry-leading long context signature + Industry-leading V2 long context。Industry-leading MIT + DeepSeek AI (Industry-leading MIT-style commercial use license + Industry-leading DeepSeek AI Liang Wenfeng founded 2023 + Industry-leading DeepSeek-Coder V2 236B MoE China code generation) で Industry-leading MIT commercial + Industry-leading Liang Wenfeng + Industry-leading V2 236B MoE。但しIndustry-leading StarCoder 2 + CodeLlama + Qwen2.5-Coder + Codestral competition (Industry-leading StarCoder 2 BigCode 619 langs 2024 + CodeLlama Meta 7B-70B 2023 + Qwen2.5-Coder Alibaba 0.5B-32B 2024 + Codestral Mistral 22B 2024 vs DeepSeek-Coder V2 236B MoE 338 langs 6T tokens 90.2% 2023 trade-off) で Industry-leading 4-code generation model competitors vs DeepSeek-Coder + Industry-leading V2 236B MoE 21B active + V2-Lite 16B 2.4B active + 338 programming languages + 6T tokens project-level + 16K-128K context + HumanEval 90.2% + MIT-style license + DeepSeek AI Liang Wenfeng 2023 unique advantage adoption alignment必須。
関連用語との違い
- vs StarCoder 2 (BigCode 2024): DeepSeek-CoderはV2 MoE 236B + 338 langs・StarCoder 2はBigCode + 3B-15B + 619 langs + dense
- vs CodeLlama (Meta 2023): DeepSeek-CoderはV2 MoE 236B + 90.2% + China・CodeLlamaは7B-70B dense + 67.8% + Meta
- vs Qwen2.5-Coder/Codestral (2024): DeepSeek-CoderはV2 MoE 236B + 6T tokens project-level・Qwen2.5/Codestralはdense
よくある質問（FAQ）
Q1: DeepSeek-Coder vs CodeLlama 違いは？ A: DeepSeek-Coder (Industry-leading 1.3B-33B V1 + V2 236B MoE 21B active + V2-Lite 16B 2.4B active + 86 languages V1 / 338 V2 + 6T tokens project-level + 16K-128K context + HumanEval pass@1 90.2% V2-Instruct 236B + MIT-style commercial license + DeepSeek AI Liang Wenfeng 2023) vs CodeLlama (Industry-leading 7B + 13B + 34B + 70B 4-tier dense + Llama 2 base + 16K-100K context + Code Llama + Python + Instruct 3 variants + HumanEval 67.8% 70B Instruct + Llama Community License + Meta AI 2023)・Industry-leading V2 MoE 236B + 338 langs + 6T tokens + 90.2% + MIT = DeepSeek-Coder + Industry-leading 7B-70B dense + Llama 2 + 100K + 67.8% + Meta = CodeLlama preference judgment。
Q2: Industry-leading V2 236B MoE + 90.2% HumanEval value は？ A: Industry-leading V2 MoE + 90.2% (Industry-leading DeepSeek-Coder V2 236B MoE 21B active + Industry-leading V2-Lite 16B 2.4B active + Industry-leading HumanEval pass@1 90.2% V2-Instruct 236B + Industry-leading V2-Instruct flagship performance)。
Q3: Industry-leading 6T tokens project-level + MIT value は？ A: Industry-leading 6T tokens + MIT (Industry-leading 6T tokens V2 project-level code training + Industry-leading DeepSeek-Coder project-level signature + Industry-leading 338 programming languages V2 + Industry-leading MIT-style commercial use license)。
まとめ
DeepSeek-Coder = 2023年DeepSeek AI Liang Wenfeng発表のChina code generation LLM。Industry-leading 1.3B + 5.7B + 6.7B + 33B V1 variants + Industry-leading DeepSeek-Coder V2 236B MoE 21B active + Industry-leading V2-Lite 16B 2.4B active + Industry-leading 86 programming languages V1 / 338 languages V2 + Industry-leading 2T tokens V1 / 6T tokens V2 project-level code training + Industry-leading 16K-128K context length + Industry-leading HumanEval pass@1 90.2% DeepSeek-Coder-V2-Instruct 236B + Industry-leading MIT-style commercial use license + Industry-leading DeepSeek AI Liang Wenfeng founded 2023 + Industry-leading DeepSeek-Coder V2 236B MoE China code generation position確立。