DeepSeek-Coder-V2-Instruct-0724
DeepSeek · 2024-07-24 · 235.7B parameters
DeepSeek-Coder-V2-Instruct-0724 is the instruction checkpoint corresponding to the DeepSeek-Coder-V2-0724 API upgrade announced July 24, 2024; DeepSeek described the update as an alignment optimization that improved general capabilities, and the downloadable Hugging Face repository was published on September 5, 2024. Its weights contain exactly 235,741,434,880 parameters, while DeepSeek reports only rounded 236B total, 21B active, and 128K context figures; the configuration declares 163,840 maximum positions. DeepSeek later combined DeepSeek-Coder-V2-0724 with DeepSeek-V2-0628 to create DeepSeek-V2.5.
Benchmark scores
| Benchmark | Score |
|---|---|
| ZeroEval CRUX | 69.5 |
| ZeroEval GSM8K | 91.5 |
| ZeroEval MMLU-Redux | 80.2 |
| ZeroEval ZebraLogic | 20.5 |