DeepSeek V4 drops 1.6T parameter MoE

DeepSeek V4 is a massive escalation in the open-weights arms race, proving that 1T+ parameter models are now the standard for frontier performance.

• –V4-Pro (1.6T) and V4-Flash (284B) directly challenge GPT-5 and Claude 4 with state-of-the-art agentic capabilities.
• –DeepSeek Sparse Attention (DSA) and KV-cache compression enable massive memory savings, making 1T+ models viable for broader inference.
• –Pricing at $3.48/M output tokens is roughly 1/7th of competitor costs, likely triggering a price war among major providers.
• –Training on Huawei Ascend 950 hardware marks a significant shift toward domestic Chinese hardware for frontier AI.
• –SOTA agentic performance on SWE-bench suggests a primary focus on deep coding and complex reasoning tasks.