Accuracy is nearly identical across Qwen3-27B NVFP4 variants, but memory footprint, inference speed, and MTP behavior differ significantly in practice.
Adapted from @Oluwaphilemon1The interesting thing about the Qwen3.8-27B NVFP4 variants is that accuracy isn’t really separating them. They’re all surprisingly close. The bigger differences show up once you actually run them. Memory footprint. Inference speed. MTP support. If you’re fighting for every bit of VRAM, minima-ai/mnma_qwen3.8_27b_nvfp4 looks like a solid option. The catch is that it doesn’t include MTP, so you lose one of the main ways to accelerate generation. NVIDIA’s NVFP4 build does have MTP, which sounds like the obvious choice if speed is the priority. But my long-context coding tests made that less convincing. MTP-4 was only around 2x faster than running without MTP. And it was still roughly 2.5x to 3x slower than the Unsloth implementation on an RTX Pro 6000. That’s a pretty significant gap. One possible explanation is how the lm_head is quantized. NVIDIA’s version uses NVFP4 for the lm_head. Unsloth keeps the lm_head in FP8. That matters because MTP relies on the target model’s predictions to accept speculative tokens. If quantizing the lm_head more aggressively makes those predictions less reliable, the acceptance rate can fall. And once acceptance drops, the theoretical speedup from MTP doesn’t translate into the same real-world throughput. So even though these Qwen3.8-27B NVFP4 variants are very close on accuracy, the inference experience can be quite different. For my setup, I’m currently leaning toward the Unsloth NVFP4 build. Not because it magically has a better base model, but because the combination of quantization choices and MTP behavior looks more attractive for actual long-context workloads. And that’s the part I find interesting about local inference. Two builds can use the same model and the same nominal precision, yet behave very differently once you start pushing context and speculative decoding. The benchmark score only tells you part of the story. The real test is what happens when you give the model a huge codebase and ask it to work for an hour. That’s what I’m testing next: long-horizon agentic coding.
