StandardThinking/StandardOne-8B-SH

Standard One 8B SH

Updated weights (v4, 2026-10-13). v4 is trained further on decision data, mostly our own synthetic items. Decision Index 0.3 public index 57.40 (self-scored with the official kit; v3 48.34). Earlier releases stay available under the tags v3, v2 (merged weights) and v1 (LoRA adapter).

Version: v4

Standard One 8B SH is Standard One 8B with a joint schema head (SH). It reads the scenario once and scores every option of every question in a single forward pass, then returns probabilities through the same POST /v1/systemone contract as Standard One. It does not generate text.

Backbone	Standard One 8B v2.2 (mistral3, Ministral 3 8B text + Pixtral vision tower), fine-tuned for the head and merged; weights in this repository
Head	joint schema head, 70.6M parameters, float32 (head/), reading the final layer and layer 25
Question types	choice (2–255 options), noul (yes/no), score (ordinal scale)
Context	up to 32,768 tokens per request
Prompt format	the chat template with no default system message (see "Prompt format")
Other formats	StandardOne-8B-SH-FP8 · StandardOne-8B-SH-GGUF (both v4)
Smaller model	StandardOne-3B-SH
License	Apache-2.0

Why a schema head

Standard One reads one answer letter per option, so a question is limited by the label alphabet and every question in a request needs its own pass over the scenario. The schema head instead attends over the hidden states of the whole request (the final layer and one intermediate layer), places every question and option in one joint graph, and returns a calibrated distribution for each question at once. In practice:

• Many options. Intent and routing questions with 100–255 options keep nearly the accuracy of small option sets.
• Many questions per scenario. All questions of a request share one pass over the scenario.
• Long scenarios. Decisions buried in long documents are found far more reliably than with Standard One v2.2.

Results

Benchmark	SH v3	SH v4 (this release)
Decision Index 0.3, public index	48.34	57.40
Areas: Knowledge / Language / Retrieval / Tools / Arts	32.1 / 51.0 / 65.8 / 60.9 / 25.5	36.9 / 57.9 / 66.4 / 84.0 / 42.5
HoVer, multi-hop claim verification (DI)	72.6	72.8
BRIGHT, reasoning-heavy retrieval over long documents, ~12k tokens per request (DI)	34.8	34.2
Decision hidden in a long document (our set, 150 items)	87.33	88.00
JevBench public: easy / original / hard (48 / 72 / 111 items)	100.00 / 95.83 / 74.77	100.00 / 94.44 / 75.68

Decision Index rows (DI) are chance-corrected skill scores (0 = random guessing, 100 = perfect) from our own runs of the public 0.3 suite with the official kit, all 140,178 scoreable requests; every benchmark of both runs is in standardone-8b-sh-v3 and standardone-8b-sh-v4. For some DI benchmarks the public train split or the benchmark's public generator was part of training (HoVer among them); evaluated items never were. The other rows are accuracy in percent, measured with the Standard One engine on identical items.

Benchmark charts

Decision Index 0.3 public index, board data of 2026-10-12; Standard One self-scored with the official kit (see the footnote of each chart). The same charts with notes: Standard One SH on the Decision Index.

Changes in v4

• Further training on synthetic decision data. v4 continues from v3 on decision items built for this round, mostly our own synthetic items (code-checked, LLM-written in varied styles), with replay of earlier data. See "Results" for the Decision Index 0.3 and the long-context checks.

Changes in v3

• Hard decision tasks. v3 is fine-tuned on about 2,200 new decision items written in the families where v2 was weakest: long policies with exceptions and overrides, multi-criteria trade-offs, judging a response against a rubric, and date and number reasoning. The items were authored by large language models, every gold answer was checked by an independent model that answered without seeing it (and recomputed where they disagreed), and items sharing text with the JevBench public set or our evaluation sets were removed. No JevBench item was used. These items are trained with softened targets (0.8 on the gold option), which keeps the head from becoming over-confident on tasks it cannot always solve. A short final pass on policy, long-document and multi-step items keeps those skills from regressing.
• Effect. The JevBench hard tier rises from 63.96 to 74.77. Confidence on that tier tracks accuracy closely (mean confidence 76.8 for 74.8 correct; expected calibration error 0.045, v2 0.097). Long-document decisions rise from 84.67 to 87.33.
• Trade-offs. The held-out decision suite is 3.0 points below v2 (87.67 vs 90.67), the hard held-out suite 4.8 points (41.17 vs 46.00) and the JevBench original tier two items (95.83 vs 98.61). Intent labels, WinoGrande and policy judgement are within half a point of v2. Keep v2 if those held-out suites matter most to you.

Prompt format

The included chat_template.jinja has an empty default system message, and the reference server renders requests without one (SH_SYSTEM_PROMPT=none, the default in code/render.py). Do not add the Ministral default system prompt: the head was trained without it, and adding it changes the hidden states it reads.

Quick start

python -m pip install --upgrade huggingface_hub
hf download StandardThinking/StandardOne-8B-SH --local-dir StandardOne-8B-SH
python -m pip install -r StandardOne-8B-SH/code/requirements.txt


python StandardOne-8B-SH/code/serve_head.py \
  --snapshot StandardOne-8B-SH --head StandardOne-8B-SH/head --port 30171

curl -s localhost:30171/v1/systemone -H 'content-type: application/json' -d '{
  "state": {"message": "Please close my card, I lost it yesterday."},
  "questions": {
    "intent": {"type": "choice", "instructions": "What does the customer want?",
               "criteria": {"cancel": "close the card", "limit": "change the limit", "other": null}},
    "urgent": {"type": "noul", "instructions": "The request is urgent."},
    "mood":   {"type": "score", "instructions": "How upset is the customer?", "criteria": ["calm", "annoyed", "angry"]}
  }
}'

The server batches concurrent requests (--max-batch-requests, --max-batch-tokens) and binds 127.0.0.1 by default. It runs on one GPU with the backbone in bfloat16 and the head in float32. On Blackwell GPUs (B200/B300) install a CUDA 13 build of PyTorch first (pip install torch --index-url https://download.pytorch.org/whl/cu130); the CUDA 12.8 build lacks kernels for those GPUs.

Fast serving with the Standard One engine

The reference server above is plain PyTorch. For production throughput and latency, serve this repository with the Standard One engine, our SGLang fork with a prefill-only decision mode and the schema-head readout (standard-one-sglang, standard-one-adapter, Apache-2.0). It is the engine we run in production. Build the image once (it replaces only the Python package of a stock SGLang image):

mkdir ctx && cd ctx
git clone https://github.com/standardthinkingai/standard-one-sglang sglang-jev
git clone https://github.com/standardthinkingai/standard-one-adapter jev-adapter
mkdir -p jit-cache/aiter jit-cache/root-cache
docker build -f sglang-jev/examples/runtime/jev/rocm/Dockerfile.decision \
  --build-arg BASE_IMAGE=lmsysorg/sglang:nightly-dev-cu13-20260929-79cafec0 -t standard-one-engine:cu13 .

Then serve the downloaded repository as it is (weights at the top level, head in head/):

docker run -d --gpus all --network host --ipc host --shm-size 32g \
  -v $PWD/StandardOne-8B-SH:/models/m:ro standard-one-engine:cu13 python3 -m sglang.launch_server \
    --model-path /models/m --decision-readout-path /models/m/head --served-model-name standardthinking/standard-one-8b-sh \
    --host 127.0.0.1 --port 30000 --tp-size 1 --model-impl sglang --dtype bfloat16 --attention-backend triton \
    --context-length 32768 --max-running-requests 64 --mem-fraction-static 0.3 \
    --chunked-prefill-size -1 --model-config-parser hf --load-format safetensors \
    --disable-radix-cache --disable-decode-cuda-graph --disable-prefill-cuda-graph --decision-deadline-seconds 110

It answers POST /v1/systemone on port 30000. Unlike the reference server, the engine requires the model field (the served model name), as in the System One API:

curl -s localhost:30000/v1/systemone -H 'content-type: application/json' -d '{
  "model": "standardthinking/standard-one-8b-sh",
  "state": {"message": "Please close my card, I lost it yesterday."},
  "questions": {"urgent": {"type": "noul", "instructions": "The request is urgent."}}
}'

Checked on one NVIDIA B300 against the reference server on 1,431 evaluation questions: 99.58 % the same answer, all 231 JevBench public answers identical. On AMD MI350X/MI355X build with BASE_IMAGE=lmsysorg/sglang-rocm:v0.5.20-rocm720-mi35x-20260929 and run with --device /dev/kfd --device /dev/dri --group-add video -e SGLANG_USE_AITER=0 instead of --gpus all.

Run locally with llama.cpp

For CPU or Apple-silicon use, follow the download, build and server instructions in StandardOne-8B-SH-GGUF. The SH head reads hidden states from the backbone, so it requires the supplied state extractor and Python runner; stock llama-server -hf alone does not execute the schema head. Once the runner is listening on port 30171, use the requests below unchanged.

Practical examples

The following invented records illustrate four integrations: receiving inventory, diagnosing an import failure, evaluating an inspection note, and checking backup evidence. They are example inputs, not benchmark results. Start either the reference server or the SH GGUF runner on port 30171 before sending them.

Warehouse receiving: reconcile a delivery record

An inventory application can ask several questions about one receiving note. Here, the model compares the recorded quantities and checks whether the paperwork supports closing the receipt.

curl --fail-with-body http://127.0.0.1:30171/v1/systemone \
  -H "Content-Type: application/json" --data-binary @- <<'JSON'
{
  "state": "Purchase order R-417 lists 48 cable reels. The receiving scan recorded 46 reels. Two pallet tags cannot be read, and the receiver has not signed the count sheet.",
  "questions": {
    "receipt_action": {
      "type": "choice",
      "instructions": "Select the next inventory step supported by the receiving note.",
      "criteria": {
        "close_receipt": "The delivered quantity is reconciled and the count sheet is signed.",
        "reconcile_count": "The recorded quantity differs from the order or pallet identification is unresolved.",
        "await_delivery": "The shipment has not reached the receiving area."
      }
    },
    "quantities_match": {
      "type": "noul",
      "instructions": "Do the receiving scan and purchase order report the same number of reels?"
    },
    "record_completeness": {
      "type": "score",
      "instructions": "Evaluate how complete the receiving evidence is.",
      "criteria": [
        "No quantity or identification evidence is available.",
        "Some evidence is recorded, but discrepancies or missing sign-off remain.",
        "Quantities and identifiers are reconciled and sign-off is recorded."
      ]
    }
  }
}
JSON

This record describes an unresolved count, unequal quantities and partial paperwork. Use those facts to inspect the returned answers. A choice question compares the descriptions you provide, a noul question returns a yes/no probability, and a score question returns an expected position on its ordered, zero-based scale. The score can lie between two levels.

Python client

Use this small standard-library client for the remaining examples. It displays the response as received so that you can examine the selected values and probability estimates.

import json
from urllib.request import Request, urlopen

def evaluate_record(record, questions):
    payload = json.dumps({"state": record, "questions": questions}).encode("utf-8")
    request = Request(
        "http://127.0.0.1:30171/v1/systemone",
        data=payload,
        headers={"Content-Type": "application/json"},
        method="POST",
    )
    with urlopen(request, timeout=300) as response:
        print(json.dumps(json.load(response), indent=2))

Data import: identify a column mismatch

An import monitor can classify a diagnostic record before selecting a repair playbook. This input contains rows, but the expected timestamp column is absent.

evaluate_record(
    "Import job J-82 read 1,200 CSV rows. Required column event_time was not found. The file contains a column named timestamp_ms. UTF-8 decoding succeeded.",
    {"failure_kind": {
        "type": "choice",
        "instructions": "Identify the import failure described in this diagnostic record.",
        "criteria": {
            "column_mismatch": "The available column names differ from the required schema.",
            "empty_input": "The file contains no data rows.",
            "text_encoding": "The importer cannot decode the file contents.",
            "unresolved": "The record does not establish a failure category."
        }
    }}
)

Equipment inspection: evaluate a filter reading

A maintenance log can map a written observation onto a defined condition scale. The criterion here is the gauge marking recorded by the inspector.

evaluate_record(
    "Inspection of ventilation unit V-12: the filter pressure gauge is above the marked replacement line. The fan guard is secured and the housing has no visible cracks.",
    {"filter_condition": {
        "type": "score",
        "instructions": "Assess the filter condition using the reported gauge position.",
        "criteria": [
            "The gauge is inside the normal operating band.",
            "The gauge is above normal but below the replacement line.",
            "The gauge has reached or exceeded the replacement line."
        ]
    }}
)

Backup verification: compare evidence with an acceptance rule

A backup catalog can check whether a verification record satisfies a supplied rule. This example requires both a matching checksum and a successful restore check.

evaluate_record(
    "Archive B-206 was written to storage. Verification reported a checksum mismatch. A restore check has not been attempted.",
    {"backup_verified": {
        "type": "noul",
        "instructions": "Does this record satisfy the rule that a verified backup must have a matching checksum AND a successful restore check?"
    }}
)

These calls only request decisions; the surrounding application decides what to do with them. Measure accuracy and choose probability thresholds using records from your own workflow. All examples send text, which is the validated input path for these releases.

Files

Path	Contents
model-0000{1..4}-of-00004.safetensors, config.json, tokenizer files	merged backbone, bfloat16 (Mistral3ForConditionalGeneration)
chat_template.jinja	chat template with an empty default system message
head/	schema_head.safetensors (float32) and schema_head_config.json (taps: final layer and layer 25)
code/	reference /v1/systemone server: request rendering, head, scoring
SHA256SUMS, release-manifest.json	checksums; source and merge details

Limitations

• Requests longer than 32,768 tokens are refused, not truncated. This is a serving limit: the backbone handles longer inputs, but the head was not trained on them.
• On our hard held-out decision suite (600 items; long policies, multi-step lookups, date and number reasoning) it scores 41.17 against 53.50 for Standard One 8B v2.2. Prefer Standard One 8B for that kind of workload.
• Text input only has been validated for this release.
• The reference server is a PyTorch server; for throughput use the Standard One engine (see above).
• Probabilities are calibrated on our held-out data; refit thresholds on your own distribution before relying on them.

License

Apache-2.0. Built on Standard One 8B (Apache-2.0), which is built on Ministral 3 8B (Apache-2.0).