This week the LocalLLaMA community dropped news: the QUASAR team released an "extreme slimmed-down" version of Qwen3-27B.
What this is
Slimming refers to quantization — compressing model parameters from high-precision numbers (BF16) down to low-precision (NVFP4, an NVIDIA-proprietary 4-bit floating-point format). The usual trade-off is accuracy loss; the benefit is a dramatically smaller model footprint and far lower GPU VRAM requirements to run it.
Key numbers:
- Original Qwen3-27B: 55.6 GB VRAM
- QUASAR compressed version: 19.7 GB (65% smaller)
- GPQA-Diamond score: 0.9141 → 0.9091 (virtually unchanged)
- AIME26 score: 1.0000 → 1.0000 (perfectly flat)
It's smaller than other NVFP4 compression efforts in the community (unsloth at 23.4 GB, Inferact at 26.4 GB), and scores higher too.
Industry view
The prevailing view is positive: VRAM dropping from 55.6 GB to 19.7 GB means a 27B model that previously needed two H100 GPUs may now run on a single NVIDIA Blackwell GPU. That's a direct tailwind for enterprise self-hosting. In the past, 4-bit quantization was seen as a "trade performance for size" compromise — this near-zero score drop shows quantization has moved past the "barely usable" stage.
But there are caveats to watch:
- Hardware lock-in: NVFP4 is exclusive to the NVIDIA Blackwell architecture — AMD, Intel, and older NVIDIA cards can't run it. Classic vendor lock-in.
- Limited test coverage: published benchmarks only cover GPQA-Diamond and AIME26. Real-world workloads (long documents, multi-turn dialogue, industry jargon) remain untested.
- Pretty paper benchmarks ≠ production stability: throughput, latency, and long-running robustness still need engineering validation.
Impact on regular people
For enterprise IT: the excuse "we can't afford GPUs to run LLMs" is being steadily hollowed out by technology. Mid-tier players like 27B now enter the "fits on one card" range, and the bar for self-hosting a private LLM keeps dropping.
For working professionals: you don't need to understand what NVFP4 is, but know this — the bill for companies self-hosting AI tools could drop by an order of magnitude. That affects your cost arguments when pushing AI projects.
For the consumer market: cheaper model serving means AI app pricing has more room to ease; meanwhile, startups depending on cloud APIs will get squeezed by self-hosting options. The competitive landscape will shift.