At the end it's not a max token problem. A 100KB text file (I already clean the vtt and extracted the plain text) should be ~20.000 - 25.000 tokens, and Llama3.1:8B manages 128K tokens.
But the text comprehension is simply wrong. I also tried gemma3:12B, same problem.
Instead Claude Sonnet 4 from Claude.ai gives me a perfect reply.