Orivael · NVFP4 precision lab · Public results
NVFP4 is NVIDIA's 4-bit floating-point format, decoded natively by Blackwell tensor cores. We quantized Llama-3.1-Nemotron-Nano-4B-v1.1 with one standard ModelOpt recipe and scored it against its own BF16 weights on 542 HumanEval+ and MBPP+ tasks, using hidden test cases, not perplexity. We tested three modes: thinking off, thinking on, and sampled.
Download the full write-up PDF · 6 pagesThinking off
NVFP4 − BF16, 95% CI ±4.5
Batch-1 speed
268 vs 141 tok/s in vLLM
Checkpoint
vs ~9 GB BF16
Best-of-8, reasoning
pooled, 3 runs, CI −2.1 to +1.4
With thinking off, NVFP4 matches BF16 on HumanEval+/MBPP+ hidden tests and runs 1.9× faster at batch 1 from a 3.7 GB checkpoint. With reasoning on, it loses 4–6 points. Best-of-8, with a selector that sees only the examples in the prompt, closes that gap: −0.4 pt pooled over three independent sampling runs on two providers.
Setup. ModelOpt 0.47.0 PTQ, NVFP4_DEFAULT_CFG, group 16, 128 wikitext
calibration samples; lm_head, embeddings and KV cache kept in BF16. Hardware: an RTX PRO 6000 Blackwell
(sm_120), served with vLLM 0.32.0 on modelopt_fp4 kernels. Scoring: EvalPlus 0.3.1 plus tests. Statistics are
paired per task: a bootstrap 95% CI and an exact McNemar p.
Greedy, up to 1,024 new tokens, plus tests across all 542 tasks.
| Pairing | b − a | 95% CI | Only a | Only b | p |
|---|---|---|---|---|---|
| BF16 r1 vs BF16 r2 | −0.7 | [−3.1, +1.8] | 26 | 22 | 0.67 |
| BF16 r1 vs NVFP4 r1 | −0.5 | [−5.0, +3.9] | 75 | 72 | 0.87 |
Same pass rate, not the same model. 147 of 542 tasks flip between the two, in near-equal numbers each way. In a later session where two BF16 runs were byte-identical, NVFP4 still flipped 135, so those flips come from precision, not scheduling noise. The 1.9× speedup is at batch 1. Fully batched, it shrinks to 1.28×.
detailed thinking on, greedy, up to 8,192 new tokens. A trace that never closes counts as no answer.
| Run | HumanEval+ | MBPP+ | All | Unclosed | Median tokens |
|---|---|---|---|---|---|
| BF16 think r1 | 0.829 | 0.645 | 0.701 | 72 | 2,022 |
| BF16 think r2 | 0.835 | 0.661 | 0.714 | 63 | 2,041 |
| NVFP4 think r1 | 0.780 | 0.606 | 0.659 | 80 | 2,431 |
| NVFP4 think r2 | 0.787 | 0.595 | 0.653 | 77 | 2,409 |
All four cross pairings favour BF16: −4.2, −4.8, −5.5 and −6.1 points, p ≤ 0.018 each. The same session's thinking-off pair landed within noise again. NVFP4 traces run about 20% longer and close less often, yet most of the tasks it loses are traces that closed with the wrong answer.
That is still no case against reasoning. Thinking is worth +21 to +23 points to both models, and NVFP4 with thinking on (0.65) beats BF16 with it off (0.47).
Model-card sampling (T 0.6, top-p 0.95) with the same seeds for both models, 8 samples per task.
A selector runs each sample against the examples written in the prompt: HumanEval docstring lines and MBPP's visible assert. The sample that passes the most examples wins. No hidden test reaches the selector. NVFP4's oracle@8 is as high as BF16's (0.832 vs 0.825): a correct answer is in its sample set just as often, it only comes up less often per sample.
NVFP4 − BF16 in points, on hidden plus tests across all 542 tasks. Replicates are disjoint best-of-k groups from sampling run 1, and at k = 8 the three independent runs; the line at k = 8 is their pooled mean.
| Run | BF16 sel@8 | NVFP4 sel@8 | NVFP4 − BF16 | 95% CI |
|---|---|---|---|---|
| 1 (Colab G4) | 0.775 | 0.788 | +1.3 | [−1.1, +3.7] |
| 2 (Brev) | 0.790 | 0.779 | −1.1 | [−3.7, +1.3] |
| 3 (Brev) | 0.792 | 0.779 | −1.3 | [−4.1, +1.1] |
| Pooled | −0.4 | [−2.1, +1.4] |
Every run sits above the −3 margin, and the pooled lower bound is −2.1. Run 1's +1.3 was the favourable draw, so the number to quote is −0.4.
| Intervention | What it did | Gap after |
|---|---|---|
| Chat-template calibration | Magicoder + UltraChat instead of wikitext | −5.2 (worse) |
| Quantization-aware distillation | Held-out KL −37%, yet 136 flips vs 135 | +1.1, n.s. |
| 16k reasoning budget | Both models slightly better | −3.7 |
| Forced answer at the cap | +4.5 to both models | −3.7 to −6.8 |
| Triad loop (rule-based critic) | +6.6 (BF16) and +7.0 (NVFP4) at the same token budget | −3.9 |
Distillation worked at what it optimises and missed what we measure. At 86% per-token agreement over ~400-token answers, almost every answer still diverges from BF16 somewhere, so the set of passing tasks just gets re-drawn. The Triad is the best single-sample method we tested. BF16 Triad best-of-4 (0.795) beats plain best-of-8 (0.785) at half the tokens. But it lifts both precisions alike, so the gap under it remains.
Why does 4-bit precision cost accuracy only when the model reasons, and is there a fix cheaper than eight samples?