Leaked DeepSeek V4 Benchmarks Reveal a Massive 1-Million Token Context Window

TL;DR AI
2 min readKey summary
Leaked benchmarks indicate DeepSeek V4 may scale from 200B to 1T parameters and could set new AI performance standards.
Leaks report DeepSeek V4 supports a 1 million token context window.
The MHC architecture enables multimodal processing of text, images, and video.
The piece mentions OpenAI Codex plugin integration for cross-platform collaboration and Microsoft multimodal research adding critique and council systems.
Enthropic's Claude Code offers enhanced 'computer use' capabilities plus session-based controls and app-specific permissions; a seven-hour DeepSeek outage raised concerns about a possible model swap.


