A new open-source tool called inside-llm enables visualization of LLM internals by loading one or two GGUF models via llama.cpp and displaying token generation in a browser interface. It streams text output with per-token probability brightness and marks divergence points between two quantizations of the same model with magenta indicators. This allows direct comparison of quantization effects at the token level.
Read original
reddit/r/LocalLLaMA