amdenterpriseai/aim-deepseek-ai-deepseek-v4-flash-0731 repository overview

⁠DeepSeek V4 Flash 0731

DeepSeek-V4-Flash-0731 is the official release of the DeepSeek-V4-Flash family from DeepSeek AI, a 284 billion total parameter mixture-of-experts (MoE) language model with 13 billion active parameters. It supports a 1 million token context window and thinking and non-thinking modes with configurable reasoning effort (low, high, max). Post-trained for agentic and coding workloads, it supports tool calling. Architecture uses hybrid Compressed Sparse Attention and Heavily Compressed Attention with Manifold-Constrained Hyper-Connections. The checkpoint is mixed precision, with MoE expert weights in FP4 and attention, norm, and router weights in FP8, and it ships a fused DSpark draft module for speculative decoding. Licensed under MIT.