Key Specifications

Benchmark Performance

Benchmark	Score	Unit	Evaluated At	Notes	Source
MMLU	72.3	%	2024-06-21	5-shot	view
HUMANEVAL	88.9	pass@1	2024-06-21	—	view
GSM8K	62.9	%	2024-06-21	0-shot CoT	view
MATH	38.2	%	2024-06-21	0-shot CoT	view
BBH	72.3	%	2024-06-21	3-shot CoT	view
GPQA	25.4	%	2024-06-21	0-shot	view
IFEVAL	56.7	%	2024-06-21	prompt_strict	view
ARC	88.5	%	2024-06-21	challenge	view
MUSR	40.3	%	2024-06-21	0-shot	view
WINOGRANDE	78.8	%	2024-06-21	0-shot	view

Name: DeepSeek Coder V2
Brand: deepseek

DeepSeek Coder V2 代码专用 MoE 模型, 128K 上下文, 支持 338 种编程语言, 代码生成能力突出。

Fornecedor	Versão	Data de lançamento	Janela de contexto	Modalidades de entrada	Modalidades de saída	Licença
Deepseek	coder-v2	2024-06-21	128K	text	text	DeepSeek License

Benchmark	Pontuação	Unidade	Notas
MMLU (Massive Multitask Language Understanding)	72.3	%	5-shot
HumanEval	88.9	pass@1	—
GSM8K (Grade School Math 8K)	62.9	%	0-shot CoT
MATH	38.2	%	0-shot CoT
BBH (BIG-Bench Hard)	72.3	%	3-shot CoT
GPQA	25.4	%	0-shot
IFEval	56.7	%	prompt_strict
ARC	88.5	%	challenge
MUSR	40.3	%	0-shot
WinoGrande	78.8	%	0-shot

Entrada	Saída	Leitura de cache	Escrita de cache
—	—	—	—

por milhão de tokens