Qwopus v3 gets NVFP4, AWQ, FP8 quants

This is less a new model than a distribution win: the interesting part is making a strong 27B-class reasoning model easier to run on real hardware. For local-LLM users, that matters more than another benchmark claim.

• –NVFP4 is the headline variant because it trims memory hardest, with the model card calling it the smallest build at about 24 GB.
• –The release is clearly tuned for vLLM users, but it is not frictionless: Blackwell needs a patched vLLM path and CUTLASS backend for NVFP4 GEMM.
• –AWQ-4bit is the compatibility play, while FP8 dynamic is the safer middle ground for teams that can spare more VRAM.
• –The model keeps the hybrid Qwen3.5 DeltaNet + softmax architecture and MTP head, so this is about preserving behavior while compressing weights, not changing the underlying model.
• –SGLang is listed as unsupported for this checkpoint, so adopters need to be aligned with the vLLM stack.