deepseek-ai/DeepSeek-V3
Read the upstream summary on the left, browse the cached forks below it, and load each fork comparison into the right-hand panel.
deepseek-ai/DeepSeek-V3
DeepSeek-V3 is a large open-source MoE language model repository from deepseek-ai with very high community interest (102,414 stars, 16,610 forks). It includes model information, weights documentation, inference code, figures, and separate code/model licenses. The README describes a 671B-parameter model with 37B activated per token, trained on 14.8T tokens, with reported strong benchmark performance and stable training.
Jump straight into Discofork's strongest cached fork picks, or open a compare view in one click.
Choose a fork to inspect
Choose upstream unless you specifically need this frozen December 2024 snapshot. This fork adds no evident capabilities and is materially behind on maintenance and support.
Choose this fork if your priority is getting DeepSeek-V3 running on a single node with lower memory use. Stick with upstream if you need the latest fixes, broader hardware support, or the most faithful reference implementation.
Prefer upstream unless you specifically need this exact June 2025 snapshot. This fork adds no visible capabilities and mainly serves as a frozen copy that lags recent fixes and platform support.
Prefer this fork if your main need is to convert, reshape, and run DeepSeek-V3 in a customized local workflow. Prefer upstream if you want the latest inference fixes, kernel updates, and broader maintenance momentum.
Prefer upstream unless you specifically need this older snapshot; this fork adds no visible value and is materially behind on maintenance and fixes.