DeepSeek Coder
#74 en Modèles de langage open sourceDeepSeek · depuis November 2023 · 9× · vu le 15 juil. 2026
17
Momentum
DeepSeek Coder is a family of eight model variants (four Base, four Instruct) released by DeepSeek in November 2023 as the company's first open-source model dedicated to coding tasks. The models were trained from scratch on 1.8 trillion tokens, of which 87% consists of source code spanning 87 programming languages. All variants support a 16K-token context window and are designed for code generation, completion, and infilling tasks. Model sizes range from 1.3 billion to 33 billion parameters.
Historique du momentum
19.05.17.08.
Fonctionnalités
| Benchmark Score (MMLU/Similar) | HumanEval: 79.3% (DeepSeek-Coder-33B-Instruct, v1); DeepSeek-Coder-V2: HumanEval 90.2%, MBPP 76.2%, LiveCodeBench 43.4% |
| Inference Speed | DeepSeek-Coder-V2 Lite (16B MoE): ~12 tokens/sec in single-stream mode; ~150 tokens/sec at batch size 28 (user-reported figures, MoE architecture favors batch processing) |
| Context Window | 16,000 tokens (16K) |
| Model Size (Parameters) | 1.3B to 33B parameters (8 variants: 1.3B, 5.7B, 6.7B, 7B, 33B; each in Base and Instruct) |
| Memory Requirement | DeepSeek-Coder-V2 Lite (16B MoE, 2.4B active): ~32 GB weights at FP16; fits on a single 16 GB GPU (e.g., RTX 4060 Ti) at INT4. 236B variant requires multiple 96 GB cards. |