DeepSeek is expanding its V4 model line with new preview access and limited-time testing of a multimodal “V4.1 Flash” model, while also updating parts of its ecosystem to default to V4. TechNode reports that V4.1 Flash enters a limited-time beta, using a new architecture with native multimodal support and positioning it as faster and lower-cost, without branding it as a formal full release. Separate coverage says DeepSeek also provides a broader V4 preview with open-source access and very long context capability up to one million tokens.
Several outlets add that DeepSeek’s V4 work is tied to multimodal capabilities, including vision and potential text, image, and video generation. In parallel, DeepSeek publications and analysis focus on cost and efficiency: Synced Review highlights a paper associated with DeepSeek-V3 training, and other reporting describes methods that reduce hardware requirements using hardware-aware co-design and sparse-attention ideas. TechNode also says OpenClaw, an open-source agent framework, updates to make DeepSeek V4 Flash its default model and compares its inference performance to a higher-parameter V4 Pro variant.
