Extended Context Window Reasoning Up To 100k Tokens

via “128k context window for extended image-text reasoning”

Mistral's 124B multimodal model with vision capabilities.

Unique: Dedicated vision encoder tokenizes images at ~4.3K tokens per image, enabling 30 high-resolution images in 128K context while maintaining text capacity, unlike models that use fixed-size embeddings or allocate disproportionate tokens to vision

vs others: 128K context with 30-image capacity exceeds GPT-4V's context window and image handling, enabling longer document analysis and more images per conversation