StandardThinking/StandardOne-35B-SH

Standard One 35B SH

Version: v1

Standard One 35B SH is Qwen3.6-35B-A3B, a mixture-of-experts model with about 3B active parameters per token, fine-tuned 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 POST /v1/systemone contract used by the other Standard One models. It does not generate text.

Backbone	Qwen3.6-35B-A3B (qwen3_5_moe, 40 layers, 256 experts with 8 active per token), fine-tuned for the head and merged; weights in this repository (bfloat16, about 70 GB)
Head	joint schema head, 68.5M parameters, float32 (head/), reading the final layer and layer 29
Question types	choice (2–255 options), noul (yes/no), score (ordinal scale)
Context	trained on requests up to 32,768 tokens
Prompt format	the Qwen chat template with thinking disabled and no system message (see "Prompt format")
Smaller models	StandardOne-8B-SH · StandardOne-3B-SH
License	Apache-2.0

Why a schema head

A generative decision model reads one answer token per option, so every question in a request needs its own pass and the option count is limited by the label alphabet. The schema head 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:

• Many options. Intent and routing questions with 100–255 options.
• Many questions per scenario. All questions of a request share one pass over the scenario.
• Long scenarios. Decisions buried in long documents.

Results

Benchmark	3B SH v4	8B SH v4	35B SH v1 (this release)
Decision Index 0.3, public index	58.08	57.40	66.26
Areas: Knowledge / Language / Retrieval / Tools / Arts	36.3 / 57.5 / 67.0 / 85.7 / 47.5	36.9 / 57.9 / 66.4 / 84.0 / 42.5	52.1 / 66.0 / 72.5 / 86.1 / 54.6
Decision hidden in a long document (our set, 150 items)	84.67	88.00	93.33
JevBench public: easy / original / hard (48 / 72 / 111 items)	100.00 / 95.83 / 63.96	100.00 / 94.44 / 75.68	100.00 / 97.22 / 78.38

Decision Index rows 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. The other rows are accuracy in percent, measured with the same engine on identical items. On the JevBench hard tier the model's mean confidence is 80.0 for 78.4 correct (expected calibration error 0.024).

Largest gains over the 8B SH (Decision Index skill points): GSM8K +41.7, BPoMP +35.1, CRUXEval +34.2, PhishNChips +32.4, GPQA Diamond +22.4, iSarcasmEval +21.4, MMLU-Pro +19.7. Intent classification, contracts and API calls stay within 1.3 points of the 8B (CLINC150, BANKING77, ContractNLI, API-Bank); RAGTruth (−2.1) and VAST (−2.2) are slightly lower.

Prompt format

Requests are rendered with the included chat_template.jinja, thinking disabled (enable_thinking=False) and no system message. The reference server in code/ does this by default. Do not add a system prompt or enable thinking: the head was trained without them, and either changes the hidden states it reads.

Quick start

python -m pip install --upgrade huggingface_hub
hf download StandardThinking/StandardOne-35B-SH --local-dir StandardOne-35B-SH
python -m pip install -r StandardOne-35B-SH/code/requirements.txt

python StandardOne-35B-SH/code/serve_head.py \
  --snapshot StandardOne-35B-SH --head StandardOne-35B-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 reference server is plain PyTorch: backbone in bfloat16, head in float32, one GPU with at least 96 GB of memory. It batches concurrent requests (--max-batch-requests, --max-batch-tokens) and binds 127.0.0.1 by default.

Limitations

• Decision Index numbers are self-scored with the official kit, not a leaderboard entry.
• Evaluated on text requests; image inputs are not covered by these results.
• Requests longer than 32,768 tokens are outside the trained range.

License

Apache-2.0 (weights and head). The base model Qwen3.6-35B-A3B is Apache-2.0.