Qwen3.5 397B wins local coding

This is less a speed win than an iteration-economy win: if the model needs fewer repair passes, the slower token rate stops mattering as much.

• –The poster’s core claim is that scale still matters for coding quality, especially bug-freeness and first-pass usefulness.
• –Concise reasoning is a big part of the appeal here; the model apparently avoids the endless-think failure mode that can make smaller reasoning models painful.
• –Running a 397B model locally at 123 GiB via IQ2_XS is notable, because it pushes “local” from hobbyist territory into serious workstation/server territory.
• –The comparison set is interesting: the biggest gains are being felt against other large open models, not just small local ones.
• –This is anecdotal, not a controlled benchmark, but it matches the broader open-weights trend toward larger MoE models becoming practical enough to compete on real developer workflows.