Strata qwen3.8flash next truly a god-tier project! 512K ultra-long context for local deployment
9321
0
2026-10-06 04:34:22
Personal opinion, for reference only
91
21
252
25
课程总结
收起全部章节
关闭课程总结
暂无课程总结
Using the Strata framework IQ3-XXS, MTP + full visual support; Project address: github.com/Niko1221/Strata Model address: www.modelscope.cn/models/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF
Computer specs: 14490f + 5060ti 16g + ddr4 64g 3200 speed averaging 40t/s; closing the visual feature can increase to around 50t/s; a 5070 graphics card can boost speed by 50%
Low power consumption, during a one-hour task, the graphics card never exceeded 60 degrees, with average power consumption of 70W;
In addition, regarding the long thinking dead loop issue, I tested Hermes for several days and found that when multiple requests and concurrent operations are made simultaneously for a task, the model will queue them in order. To avoid wasting time due to queue timeouts, I had the AI switch to single-request serial operations and have not encountered any dead loops since.
Comparing the original version with Swift 1.5:
Original Strata: success rate for long tasks is much higher than Swift 1.5, often completing in one go, but the high thinking cost makes it time-consuming.
Swift 1.5: thinking time is reduced by ≈46% compared to the original, but long tasks may require rework.
