---
base_model: StandardThinking/StandardOne-8B
library_name: transformers
license: apache-2.0
pipeline_tag: text-generation
tags:
  - mistral3
  - decision-model
  - typed-decisions
  - jev
  - jevbench
  - fp8
  - compressed-tensors
  - vision-language
---

# StandardOne-8B-FP8

**Version:** v2.2

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 `2b31eebec5ce6aa4c023c8e19f1af9b2e2688200` of StandardOne-8B (v2.2) on 2026-10-04 using
llm-compressor 0.14.0 (torch 2.14.0, transformers 5.17.0,
compressed-tensors 0.19.0); results below.

## Validation

Both precisions were served the same way through SGLang 0.5.20 and `jev-adapter` (`served` wording, one option order, default temperature 1.65), measured 2026-10-04. Accuracy is the most probable answer and does not depend on temperature.

| Suite | BF16 (StandardOne-8B v2.2) | FP8 (this repository) | Change (points) |
|---|---:|---:|---:|
| many-option questions, 53–151 options (18,000) | 82.67 % | 82.34 % | −0.33 |
| the same question set, at most 26 options (750) | 88.13 % | 88.13 % | 0.00 |
| long-document questions (150) | 40.00 % | 40.67 % | +0.67 |
| held-out decision set (600) | 82.50 % | 84.00 % | +1.50 |
| hard proxy (600) | 53.50 % | 52.67 % | −0.83 |
| realistic transfer set (600) | 91.50 % | 90.83 % | −0.67 |
| JevBench public easy (48) | 100.00 % | 100.00 % | 0.00 |
| JevBench public standard (72) | 98.61 % | 97.22 % | −1.39 |
| JevBench public hard (111) | 56.76 % | 55.86 % | −0.90 |

Across 28,406 validation questions with recorded probabilities, FP8 and BF16 gave the same answer for 95.83 %. No separate temperature was fitted for this build; use the serving settings of [StandardOne-8B](https://huggingface.co/StandardThinking/StandardOne-8B). Score changes from v2.1 to v2.2 are listed in the [StandardOne-8B card](https://huggingface.co/StandardThinking/StandardOne-8B#changes-in-v22).