Laguna 2.1 solved a hard memory-constrained coding task after 60k+ thinking tokens

pand5461 · reddit · 2026-07-25

A Reddit user says Laguna 2.1, a 120B-class model, solved a hard constrained-memory data rearrangement problem that local Qwen models failed on.

Key points from the test:

The task required converting a Union-Find-like id array into a compact cluster list structure.
The solution had to work within a fixed memory budget, with no dynamic allocations.
Laguna produced more than 60k thinking tokens before generating code that passed the tests.
The code worked, but used a hack: packing two smaller integers into one 64-bit value, which could fail in edge cases.

The author argues that this kind of very long reasoning may be awkward for ordinary coding tasks, but useful for hard debugging and review work.

More from coding & agent