FlagRelease/DeepSeek-V4-Flash-hygon-FlagOS · Hugging Face

# Introduction DeepSeek-V4-Flash is one of two models in the V4 series released by DeepSeek. It uses a Mixture of Experts (MoE) architecture with 284B total parameters, only 13B of which are activated, and supports a context length of up to 1 million tokens. Architecturally, the model introduces a hybrid attention mechanism, manifold-constrained hyperconnections, and the Muon optimizer. Pre-training data exceeds 32T tokens, and post-training follows a two-stage paradigm — first independently cultivating domain experts via SFT and GRPO reinforcement learning, then unifying multi-domain capabilities into a single model through online policy distillation. In maximum reasoning mode, a larger thinking budget allows its reasoning performance to approach that of the Pro version; however, due to its smaller parameter scale, it falls slightly short of Pro on pure knowledge tasks and the most complex agent workflows. ### Integrated Deployment - Out-of-the-box inference scripts with pre-configured hardware and software parameters - Released **FlagOS-Hygon** container image supporting deployment within minutes ### Consistency Validation - Rigorously evaluated through benchmark testing: Performance and results from the FlagOS software stack are compared against native stacks on multiple public. # Evaluation Results ## Benchmark Result | Metrics | DeepSeek-V4-Flash-Nvidia-Origin | DeepSeek-V4-Flash-Hygon-FlagOS | |--------------|--------------------------------|--------------------------------------| | GPQA | 0.697 | 0.702 | | Aime | 0.7333 | 0.7 | # User Guide Environment Setup | Item | Version | |------------------|----------------------| | Docker Version | Docker version 20.10.24, build 297e128 | | Operating System | Sugon OS 8.9 | ## Operation Steps ### Download FlagOS Image ```bash docker pull harbor.baai.ac.cn/flagrelease-public/flagrelease-hygon-deepseek-v4-flash:202604242039 ``` ### Download Open-source Model Weights ```bash pip install modelscope modelscope download --model FlagRelease/DeepSeek-V4-Flash-hygon-FlagOS --local_dir /data/DeepSeek-V4-Flash-hygon-FlagOS ``` ### Start the Container ```bash export NCCL_SOCKET_IFNAME=eno1 export NCCL_IB_DISABLE=1 export NCCL_DEBUG=INFO export USE_FLAGGEMS=1 export NCCL_DEBUG=WARN export GEMS_VENDOR=hygon docker run \ --name flagos \ --network=host \ --ipc=host \ --device=/dev/kfd \ --device=/dev/mkfd \ --...

Source context: 0 downloads · 0 likes · Repo FlagRelease/DeepSeek-V4-Flash-hygon-FlagOS

Source context: 0 downloads · 0 likes · Repo FlagRelease/DeepSeek-V4-Flash-hygon-FlagOS