NVIDIA has demonstrated that NVFP4 can achieve inference without loss of precision, with significant performance and energy-efficiency gains, and is equally suitable for training.

Judged as of 2026-03-17. The claim has three parts: (a) NVIDIA has published or presented results showing NVFP4 inference with no meaningful loss of accuracy against a higher-precision baseline such as FP8 or BF16 ("without loss of precision" means accuracy on standard benchmarks is the same or within about 1%, not bit-identical); (b) significant gains in throughput or performance and in energy efficiency over higher-precision formats; (c) results showing NVFP4 training (for example a pretraining run that matches an FP8 or BF16 baseline). True: published results by that date (technical blog, paper, MLPerf submission or independent replication) show all three. Mostly true: all three are shown, but with a small slip, such as accuracy loss slightly above 1% on some models, or training shown only at limited scale. Mixed: one of the three parts is unsupported, or the evidence shows clear accuracy drops on many models. Mostly false: only performance gains are shown, and the accuracy and training claims rest on narrow or early results that are stated as general. False: the evidence shows substantial accuracy loss in NVFP4 inference, or no NVFP4 training results existed by that date.

Set before the research began.

1. E1 NVFP4 Trains with Precision of 16-Bit and Speed and Efficiency of 4-Bit | NVIDIA Technical Blog NVIDIA Technical Blog (NVIDIA) · 25 August 2025 Own record

   “Figure 4 illustrates that across all domains, NVFP4 matches the performance of FP8, highlighting its effectiveness.”

2. E2 Paper page - Pretraining Large Language Models with NVFP4 Hugging Face (paper page for arXiv 2509.25149) · 30 September 2025 Own record

   “Our results show that the model trained with our NVFP4-based pretraining technique achieves training loss and downstream task accuracies comparable to an FP8 baseline.”

3. E3 NVIDIA Blackwell Ultra Sets New Inference Records in MLPerf Debut | NVIDIA Technical Blog NVIDIA Technical Blog (NVIDIA) · 9 September 2025 Own record

   “This optimization led to reduced model size and the ability to use the higher-throughput NVFP4 compute built into Blackwell and the even higher throughput in Blackwell Ultra—all while meeting the strict target accuracy of the benchmark.”

4. E4 NVFP4 Trains with Precision of 16-Bit and Speed and Efficiency of 4-Bit | NVIDIA Technical Blog NVIDIA Technical Blog (NVIDIA) · 25 August 2025 Own record

   “NVFP4 training is still in the research phase, exploring and validating the potential of 4-bit precision in large-scale model pretraining.”

On training, NVIDIA's blog says NVFP4 matches FP8 across all domains (E1), and its pretraining paper reports training loss and downstream accuracy comparable to an FP8 baseline (E2). NVIDIA also notes that NVFP4 training is still in the research phase (E4). On inference, NVIDIA's MLPerf write-up says an optimisation let it use NVFP4's higher-throughput compute while still meeting the benchmark's strict target accuracy (E3). That gives some support for accuracy and throughput, but there is no direct comparison with FP8 or BF16 inference and no energy-efficiency figures. Since the energy-efficiency part is unsupported and the inference-accuracy part is only partly supported, the claim resolves as mixed. All of this evidence comes from NVIDIA itself.