---
license: cc-by-4.0
tags: [benchmarks, quantization, nvfp4, mtp, llm-serving]
---

# protoLabs lab-benchmarks

Every number on a protoLabsAI model card traces to a row here. Release-gate results
(quant vs bf16 baseline, paired task sets, outliers re-trialed x3 both sides), speed-test-v2
regime matrices (InferenceMAX-style: seeded random dataset, client-side TTFT/TPOT p50/p99,
goodput), decode-at-depth ladders, and coherence-probe verdicts.

Methodology: single-stream-only numbers are never published without load numbers;
spec-decode accept% is reported per workload (random-data benches understate it ~2.5x);
LLM-judged suites are gated behind deterministic ones. Harness: protoLabsAI/protoLab evals.

`vram_gb` on each row is the artifact's published weight-file size in GB (the download you
load) — the traceable X-axis for the quality-vs-VRAM chart. GGUF rows use the specific quant
variant's file (e.g. the NVFP4 gguf, not a Q8_0 in the same repo), not the repo total.

CC-BY-4.0 — cite, steal, argue. Charts: protolabs.studio/lab