ByteDance Research found that DeepSeek-V4 in long-text processing, answers periodically change as the length of irrelevant content varies, while V4.1-Flash the fluctuations narrow somewhat but persist.
The report says Apple open-sourced a Qwen3.5-9B-Base trained on LensVLM-9B, which processes long documents by filtering with low-resolution pages before reading the original text. It says this can significantly reduce context and KV cache usage, but each answer takes longer.
