I have been meaning to test out the Flint model from Recursive Superintelligence on my Rust program in Pi code. Flint was released last month, and has impressive benchmarks (it outperforms Laguna XS 2.1 and Qwen 3.6 35B). It is also fast as it only uses 2.5 billion active parameters (of 40 billion total).
After some issues (aren't there always issues) I got Pi working with llama-cpp on the local model router. I had Flint remove all unsafe code in my Rust program and add the forbid_unsafe directive. It took 30,000 tokens and proceeded at a nice pace (40 tokens per second), and my PC handled the load easily.
It correctly made the required changes and even used the utility functions specific to my project. So I was impressed, both on Flint's speed and the quality of its output. The Rust program compiled correctly. Unfortunately, the version with no unsafe code is much slower, so I discarded it—but this was not the fault of the LLM, it was just an experiment.
I am inclined to believe in Recursive's benchmarks: Flint (Flint-Q4_K_M.gguf) is probably a stronger coding model than the Laguna or Qwen models it was compared against. And this means it is likely my first choice for local LLM coding now.