Swift1.5-Qwen3.8-Flash-Next benchmarked: 40% of tokens and time, near-identical quality

returnity · reddit · 2026-09-26

Reddit user returnity published a detailed Aider agentic coding benchmark comparing UkisAI's Swift1.5-Qwen3.8-Flash-Next against base Qwen3.8-Flash-Next (both Q5 quants, 262k context, 128GB setup).

Headline result: quality is statistically indistinguishable, efficiency is dramatically better.

First-try pass: 41.1% vs 40.2%; retry pass 86.9% vs 90.7% (McNemar p≈0.29, not significant)
Swift uses just 40% of median tokens (6,991 vs 17,646) and finishes in 40% of the wall-clock time (608 vs 1,542 sec/case)
Well-formed diff rate: 100% vs 98.1%

Key insight: the base model frequently goes on long reasoning binges, looping back on itself; Swift almost never does. On the base model's 20 most token-hungry runs, Swift used 29% of the tokens and solved 16/20 vs 17/20; max tokens per case were 44k vs 203k.

By language: C++ sees the strongest compression (29% of tokens) but also 3 of 6 losses; Python/JS sit around 46%. The author credits UkisAI's RL/OPD training and calls it the best local model he has used.

Original post →

Swift1.5-Qwen3.8-Flash-Next benchmarked: 40% of tokens and time, near-identical quality

More from coding & agent