How does DeepSeek V4 compare to GPT-5.4?

V4 matches O3-level reasoning at 1/20th the cost, with native multimodal support including video/audio generation.

What does 1 trillion parameters mean?

Comparable to GPT-4's ~1.8T parameters but with significantly improved architecture efficiency.