Recursive llama.cpp Engine on RTX 3050: Qwen 32B and 7B Coder Tests — 3x Quality Boost and Memory Appetites

The author of the article continues testing a fork of llama.cpp that implements a recursive mode (designated D=12) against the standard engine (D=0). The tests were run on an RTX 3050 6 GB video card, with all models in Q4 quantization. For the large Qwen 32B-A3B model, on easy benchmark tasks both engines produced identical results, but the recursive version was slightly slower. On the hard segment, the recursive engine fixed almost 3.1 times more errors in the code review task, although it took 17% more time. The author notes that the recursive engine solves the problem of 'lazy code generation', producing complete monolithic scripts instead of fragmented snippets. For Qwen 2.5 Coder 7B Instruct, the recursive version performed better in code review (found one more bug), in fullstack development it fully satisfied the requirements while the standard version did not, and in the 2D game test it was a draw, with the recursive engine avoiding non-existent tags and implementing basic game logic. For Mistral 7B Instruct v0.2, the recursive engine significantly improved security (98% protected), dramatically improved the fullstack platform (2.2x improvement, 100% business logic), and produced cleaner game layout. The main drawback is a speed reduction of about 13% and increased VRAM consumption. The author links the fork repository and Google Drive with detailed results.