Qwen 3.8 Flash Next Q3 S - Benchmark creation
retained text
After running a benchmark test on Exo in Opencode and being amazed by the result, I thought I'd try other models on Opencode. Many were surprisingly terrible. Especially some mid level models. Earlier posts: [https://www.reddit.com/r/opencode/comments/1wzm1ah/exo\_free\_frontier\_level\_wow/](https://www.reddit.com/r/opencode/comments/1wzm1ah/exo_free_frontier_level_wow/) [https://www.reddit.com/r/opencode/comments/1ww5o7y/animated\_horse\_space\_bunny/](https://www.reddit.com/r/opencode/comments/1ww5o7y/animated_horse_space_bunny/) Today I installed Strata on [https://github.com/Niko1221/Strata](https://github.com/Niko1221/Strata) I connected up my AMD R9700 32gb with 64gb RAM and used the model **Qwen3.8-Flash-Next IQ3\_S**. Admittedly I was expecting terrible results considering it's a small Q3. It acheived 90 tk/s decode approximately 3x compare to 27b. I tried the usual prompt from the other posts, but the output wasn't great. So I tried iterating with a few more prompts in the hope it would improve it. The major leap was when I asked it for a plan, then asked it to build it and when I said "low poly" style but with many polygons. FYI I'd done similar with other models and the result would get worse rather than better. Anyway I thought I'd share the output. **Qwen 3.8 Flash Next IQ3** **S -** [**https://web.esuyo.com/horse-strata-qwen-3.8-q3-s-1/**](https://web.esuyo.com/horse-strata-qwen-3.8-q3-s-1/) Qwen 3.8 Max - [https://web.esuyo.com/horse-qwen-3.8-max/](https://web.esuyo.com/horse-qwen-3.8-max/)
document fetched 2026-10-13 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1x4h1c2/qwen_38_flash_next_q3_s_benchmark_creation