DeepSeek-Coder-V2-Lite-Base
DeepSeek · 2024-06-14 · 15.7B parameters
DeepSeek-Coder-V2-Lite-Base is the uninstructed base checkpoint of DeepSeek's open-weight MoE code model. Its downloadable weights contain exactly 15,706,484,224 parameters, while DeepSeek reports rounded figures of 16B total and 2.4B activated per token; DeepSeek documents 128K context support, although the YaRN configuration declares 163,840 maximum positions. It was continued-pretrained from an intermediate DeepSeek-V2 checkpoint, bringing reported cumulative exposure to 10.2T tokens (4.2T from the DeepSeek-V2 data plus 6T additional tokens), and supports 338 programming languages.
Benchmark scores
| Benchmark | Score |
|---|---|
| ARC AI2 | 57.3 |
| BigCodeBench Complete | 30.6 |
| BigCodeBench-Hard Complete | 11.5 |
| Capability | 102.9 |
| Epoch Capabilities Index (ECI) | 108 |
| GSM8K | 67.1 |
| MMLU | 60.5 |
| WinoGrande | 72.9 |