Llama.cpp b9870 Release Fixes Long Reasoning Loops

How it developed

1. July 4, 2026llama.cpp ReleasesWhere it started

   Llama.cpp b9870 Release Fixes Long Reasoning Loops

2. July 13, 2026llama.cpp Releases

   Llama.cpp Update Fixes Reasoning Budget Handling

   Fixes per-request reasoning budget handling by ensuring caller-supplied values reach the sampling layer instead of defaulting to server sett

3. July 15, 2026llama.cpp Releases

   llama.cpp b10003 release enhances tokenization

   llama.cpp b10003 enhances tokenization via unified argument parsing, Windows UTF-8 support, and exposed model-sourcing flags