---
base_model: StandardThinking/StandardOne-3B-SH
base_model_relation: quantized
library_name: transformers
license: apache-2.0
pipeline_tag: text-classification
tags:
  - mistral3
  - decision-model
  - schema-head
  - fp8
  - compressed-tensors
---

# Standard One 3B SH — FP8

> **Updated weights (v2, 2026-10-11).** This repository now holds the FP8 build of
> [Standard One 3B SH](https://huggingface.co/StandardThinking/StandardOne-3B-SH) v3. The previous build stays
> available under the tag `v1`.

**Version:** v3 (2026-10-13)

An FP8 (compressed-tensors, float8_e4m3 weights, dynamic per-token activations) quantization of the merged backbone
of [Standard One 3B SH](https://huggingface.co/StandardThinking/StandardOne-3B-SH) v3, with the same schema head.
The language-model linear projections (q/k/v/o, gate/up/down; 182 tensors) are quantized per output channel
(scale = absmax / 448); the vision tower, multi-modal projector, embeddings, norms and lm_head stay in BF16. This is
the same data-free recipe and `quantization_config` as
[StandardOne-3B-FP8](https://huggingface.co/StandardThinking/StandardOne-3B-FP8). The head (`head/`) is float32 and
identical to the BF16 release.

| | |
|---|---|
| Weights | 4.4 GB (BF16 release: 7.7 GB) |
| Head | `head/schema_head.safetensors`, float32, taps: final layer and layer 19 |
| Prompt format | chat template with no default system message (see the [main card](https://huggingface.co/StandardThinking/StandardOne-3B-SH#prompt-format)) |

## Serving

Serve it with the **Standard One engine** (our SGLang fork with the schema-head readout; build and run commands in the
[main card](https://huggingface.co/StandardThinking/StandardOne-3B-SH#fast-serving-with-the-standard-one-engine)).
Use this repository as `--model-path` and its `head/` as `--decision-readout-path`; the other arguments are the same.
Engines that load compressed-tensors FP8 read the backbone; the schema head needs the Standard One engine or the
PyTorch reference server. The reference server in StandardOne-3B-SH is written for the BF16 checkpoint; use the BF16
release with it.

## Validation

Measured 2026-10-13 on one AMD Instinct MI350X with our SGLang-based decision engine (ROCm build, with the fix in the notes),
against the official Decision Index 0.3 run of the BF16 release (StandardOne-3B-SH v3) on the same requests. The sample takes 150
random requests from each of six benchmarks where the model is strong (Home appliance simulator, ContractNLI, HoVer, CLINC150, BANKING77, POP909-CL) and six where it is weak
(MuSR, WinoGrande, GSM8K, CRUXEval, ChessBench, MMLU-Pro); ContractNLI has 123 in total. Accuracy is the most probable answer against the benchmark's gold.

| Benchmarks | Questions | BF16 | FP8 (this repository) | Change (points) | Same answer as BF16 |
|---|---:|---:|---:|---:|---:|
| strong (6 benchmarks) | 4,209 | 92.68 % | 92.63 % | -0.05 | 99.31 % |
| weak (6 benchmarks) | 810 | 46.42 % | 46.79 % | +0.37 | 94.44 % |
| all | 5,019 | 85.22 % | 85.24 % | +0.02 | 98.53 % |

Answers change almost only where the model is unsure: on the strong benchmarks FP8 gives the same answer as BF16 for 99.3 % of questions, on the weak ones for 94.4 %, and accuracy stays within a fraction of a point in both. Per-benchmark numbers are in `release-manifest.json`.

## Notes

- On AMD ROCm with the native per-token activation-quantization fallback, SGLang 0.5.20 needs a small fix when the
  engine pads the token dimension (the per-token scale buffer is padded but the fallback writes unpadded rows);
  without it the server fails at warmup. This release was validated on NVIDIA (CUDA), where the fallback is not used.

## License

Apache-2.0, the same terms as [StandardOne-3B-SH](https://huggingface.co/StandardThinking/StandardOne-3B-SH),
built on [Standard One 3B](https://huggingface.co/StandardThinking/StandardOne-3B) and
[Ministral 3 3B](https://huggingface.co/mistralai/Ministral-3-3B-Instruct-2512) (Apache-2.0).