Qwen3.8-27B NVFP4 Delivers 2.5× Speed

The compelling takeaway is not simply that four-bit inference is faster; it is that a workload-specific quality check erased BF16’s primary advantage. NVFP4 now looks like a practical capacity-and-throughput strategy for local AI developers, though one document benchmark cannot establish universal parity.

• –Unsloth lists NVFP4 at about 2.5× faster than BF16, with estimated memory falling from 56GB to 17–19GB for Qwen3.8-27B. [Unsloth](https://unsloth.ai/models/qwen3.8-27b)
• –The result is especially relevant for GX10, DGX Spark, RTX 50-series, and other Blackwell-class systems with native FP4 acceleration.
• –BF16 remains the safer baseline for sensitive evaluations, long-tail prompts, fine-tuning, and workloads where small numerical differences can affect correctness.
• –Developers should reproduce the comparison with fixed prompts, decoding settings, context lengths, kernels, and speculative-decoding configurations before generalizing the result.
• –Qwen3.8-27B became available on August 14, making this an early real-world inference benchmark rather than a model-launch announcement. [Qwen](https://github.com/QwenLM/Qwen3.8)