Qwen3.8 Flash Next - Templates Comparison

I was running into a lot of posts that praised both the Fixed template and the Sharp template in comparison to the stock one, so I put them to the test. It's not as extensive as it should be for a paper-grade analysis, but it gives out the point of each template. Test setup I used SWE-bench Verified with mini-SWE-agent 2.4.6 , slice 0:100 (the identical 100 tasks for all runs) Hardware CPU: Ryzen 9 9900X RAM: 128 GB DDR5-5600 GPU: RTX PRO 6000 WS Runtime I containerized jpezzulli/sglang-rtxpro6000 and ran Flash Next with RadixArk/Qwen3.8-Flash-Next-NVFP4 on CUDA 13.3 . Full 262K context BF16 KV 51.2 GB FP8 n-gram embedding table pinned in RAM 32 GB HiCache pinned in RAM I ran all templates at both medium and xhigh reasoning efforts. Results Metric Stock (medium) Stock (xhigh) Stock Δ Fixed (medium) Fixed (xhigh) Fixed Δ Sharp (medium) Sharp (xhigh) Sharp Δ Resolved 91 99 +8 87 98 +11 94 94 +0 Resolution rate 91% 99% +8 pts 87% 98% +11 pts 94% 94% +0 pts Median output tokens 5,691 13,855 +143.5% 6,956 14,819 +113.0% 8,596 12,008 +39.7% Median reasoning tokens 3,050 8,759 +187.2% 3,809 9,063 +137.9% 5,437 7,967 +46.5% Median wall time 38s 1m 46s +180.4% 43s 1m 47s +152.3% 1m 1m 32s +53.4% Total wall time 1h 47m 1s 4h 31m 22s +153.6% 1h 59m 53s 4h 4m 52s +104.3% 2h 29m 18s 3h 11m 36s +28.3% preview.redd.it/02geu81o8qnh1.png preview.redd.it/v2mt6mgo8qnh1.png preview.redd.it/ph1z36zo8qnh1.png Takeaways preview.redd.it/6ydu12mp8qnh1.png Raising reasoning effort to xhigh closes almost all of stock's and fixed's gap to Sharp. At medium, Sharp led resolution by +3 tasks over stock and +7 over fixed ; at xhigh, stock and fixed instead lead Sharp by +5 and +4 tasks , respectively. Sharp barely moves on resolution ( 94 → 94 ) despite a real token/time cost increase, median reasoning tokens rise +46.5% and median wall time +53.4% . This suggests it was already extracting most of the benefit it could get from extra reasoning budget at medium, while stock and fixed still had headroom. Sharp remains the most token-efficient per resolved task at xhigh ( 14,541 output tokens/resolved vs. ~17,000 for stock/fixed ), consistent with its medium-era efficiency edge, but it's no longer the highest-resolving template once reasoning effort is high. Absolute cost scales heavily with reasoning effort: total wall time roughly 2.3–2.5× for stock/fixed and +28% for Sharp; total reasoning tokens roughly doubled for stock/fixed and increased +35% for Sharp. Conclusion Sharp should be used at medium and it keeps a reasonable accuracy at very good speed. I don't see the point in using it at xhigh. By sacrificing a small accuracy you complete the tasks in half the time. Stock is the slowest but the most precise. Fixed is the middle ground between Stock and Sharp both in accuracy and speed The next benchmark will be on a much extensive SWE-bench Multilingual + Terminal Bench. Disclaimer: I wrote the post myself then used AI to format it properly for readability