Blackwellboy/DeepSeek-V4.1-Flash-1017-BLACKFOLD-TP2
Uncensored DeepSeek-V4.1-Flash-1017 on 2x DGX Spark: refusals 94 to 6, same answers as DeepSeek's API, 2,008 tok/s prefill at 128K, 113 tok/s code decode. TensorFold native, 3 engine patches, 2 RoCE rails.
GitHub repository with 38 stars and 1 forks.
Language: Python