The last task of the day is Qwen 3.8 Flash next Coder set to prefill 2K Tok/s and decode 90 Tok/s.
I am testing Pagoda as a model. In fact, I am somewhat excited because the reporting performance is high compared to the quantization level. If the performance is good and the DRAM is sufficient, for those who were waiting for the existing 35B
Make a medieval castle from real Lego bricks, and then provide all the steps and blocks I need to build it.
Qwen 3.8 27b, locally on my laptop. I'm not really impressed with the architectural choices yet, but the instructions are clear and you
I know it's hard to believe, but this IQ_3 of Qwen 3.8 Flash Next is genuinely impressive.
Sure, it's not BF16, but it works very well in Claude Code. To eliminate the likelihood of benchmaxxing, I present a novel voxel benchmark: Voxel Sermon on the Mount. This was generated
GPT-6.1 Sol doesn’t need to write the code.
It can just be the brain. Sol plans + reviews. Qwen 3.8 27B writes locally. $0.75 → $0.17 API cost for 3 small games. That’s a 77% cut. The trade-off? 6.6 min → 43.4 min. Cloud brain. Local hands.
You have reached the end of the archive
All of qwen38