StandardOne-8B-FP8 is an FP8 (compressed-tensors, float8_e4m3 weights, dynamic per-token activations)
quantization of the released StandardOne-8B decision model. The language-model linear
projections (q/k/v/o, gate/up/down) are quantized per-channel FP8 E4M3 with dynamic FP8
activations (llm-compressor's data-free FP8_DYNAMIC recipe, no calibration data required);
the vision tower, multi-modal projector, embeddings and lm_head are left unquantized in BF16.
It was produced from source revision e88423700bb5ab9b2f50e176cf19825914345272 of StandardOne-8B on 2026-09-25 using
llm-compressor 0.14.0 (torch 2.14.0, transformers 5.17.0,
compressed-tensors 0.19.0); results below.
Validation
Served through SGLang 0.5.20 and jev-adapter (native wording, no system prompt, one option order) on the
same items as the BF16 release, measured 2026-09-25. Accuracy is argmax and does not depend on temperature.
| Suite |
BF16 (StandardOne-8B) |
FP8 (this repository) |
| JevBench public easy (48) |
100.00 % |
100.00 % |
| JevBench public standard (72) |
94.44 % |
94.44 % |
| JevBench public hard (111) |
54.95 % |
51.35 % |
| judge proxy (600) |
90.50 % |
90.17 % |
| realistic transfer set (600) |
90.33 % |
90.00 % |
| stated-distribution probability (1,036) |
81.37 % |
81.27 % |
| hard proxy (600) |
52.50 % |
52.33 % |
Temperature refit on this checkpoint's own served probabilities (same held-out calibration data as the BF16
release): T = 1.4 (BF16 release: T = 1.65). Serve with --default-temperature 1.4.
