A 671B-parameter mixture-of-experts model trained with multi-token prediction and FP8 precision, DeepSeek's flagship non-reasoning chat model released in December 2024.

Benchmarks

What its maker published, and what anyone else measured. Every figure links the document it came from.

benchmark	published	measured
MMLU-Pro	75.9	—
GPQA	59.1	—
SWE-bench	42	—

Lineage

↓ finetune DeepSeek V3.1

view in full graph →

Metadata

org	DeepSeek
released	2024-12-26
params	671B-A37B
license	mit
hf id	deepseek-ai/DeepSeek-V3