build1 publisher
NVFP4 squeezes Qwen3.8's 2.4 trillion weights onto eight B300s at 150 GB a GPU
Alibaba's largest open-weight release fits on a single eight-GPU node only because a community four-bit build takes it down to 1.2 TB. AWS publishes the vLLM config for it and skips the price.
Publishers:aws.amazon.com
Reality
- Evidence52
- Adoption22
- Hype gap+32
- Incentives78
- Confidence48