DeepSeek V4 ships with 1M-token context and open weights as valuation jumps to $45B

DeepSeek released V4 in 'Pro' (1.6T total params) and 'Flash' (284B) variants with a 1M-token context window and open weights, optimized for agentic coding, research and document analysis. FT and Bloomberg report DeepSeek is in talks for its first VC round at a $45B valuation — up from $20B just weeks ago — and V4 was tuned for inference on Huawei's 950PR chip even though training still used NVIDIA.