Finally tested Alibaba_Qwen Qwen 3.8 27B on two RTX 5090 GPUs.
🔥 Getting around 120–130 tokens/sec with vLLM, UnslothAI NVFP4, native MTP, and a 150K context target. The coding output looks very good.
After deploying Qwen 3.8-27B locally, I learned another knowledge
Dgx spark is not suitable for running Dense models Different model architectures have different throughput requirements for underlying computing power and memory bandwidth. If a dense model does not do quantification and KV management well, no matter how powerful the hardware is, it will not work. Sure enough, practice brings true knowledge, let’s feel the speed.
QWEN 3.8:27b made this as a one shot prompt for me in 20m!
All while running locally on my machine. Kinda crazy
Agentic Task: Install K3s in your own slicervm
Left: DeepSeek V4 Flash 0731 - 2x DGX Sparks Right: Qwen 3.8 27B (FP8) - 1x RTX 6000 Pro 40 (ish) tok/s gen vs 88 tok/s Why do they both feel "slow"? Thinking aka "variant" - they're spending a lot of tokens on reasoning.
You have reached the end of the archive
All of qwen38