InferCrane Commerce-1

A 27B, self-hosted decision model for commerce agents. Give Commerce-1 one structured state and ordered Choice, Noul, or Score questions; it returns calibrated probability distributions and typed answers—not prose—through an offline, content-verified runtime.

Canonical repository: infercrane/Commerce-1
Runtime and examples: github.com/infercrane/commerce-1

What it returns

Primitive	You provide	Commerce-1 returns
Choice	Named candidate actions and their criteria	Selected action plus a probability for every action
Noul	One yes/no decision instruction	Boolean answer plus P(true)
Score	Ordered labels from low to high	Expected ordinal score plus a probability for every label

The API also returns the raw ordered distributions, immutable model and runtime identities, and measured input-token use. Probabilities help callers set review thresholds, but they do not guarantee correctness or transfer unchanged to a new workload.

Intended use

Commerce-1 is designed for structured decisions inside commerce-agent workflows, including proceed/review/block gates, tool routing, approval-policy checks, action selection, and calibrated risk or priority scoring.

It is not a chat model, search engine, live-fact retriever, payment processor, or autonomous authority for refunds, disputes, fraud, credit, safety, or legal decisions. Validate on representative traffic and retain explicit controls for consequential actions.

Run locally

Use immutable 40-character revisions from the verified release receipts. Floating source branches and model revisions are not part of the qualified path.

export COMMERCE_ONE_SOURCE_REVISION=<verified-github-revision>
export COMMERCE_ONE_MODEL_REVISION=<verified-huggingface-revision>

git clone https://github.com/infercrane/commerce-1.git
cd commerce-1
git checkout --detach "$COMMERCE_ONE_SOURCE_REVISION"
test "$(git rev-parse HEAD)" = "$COMMERCE_ONE_SOURCE_REVISION"

uv sync --frozen --python 3.12 --extra v3-runtime
uv run --frozen --python 3.12 --extra v3-runtime hf download infercrane/Commerce-1 \
  --revision "$COMMERCE_ONE_MODEL_REVISION" \
  --local-dir ../commerce-1-model

export COMMERCE_ONE_MODEL_DIR="$(cd ../commerce-1-model && pwd)"
uv run --frozen --python 3.12 --extra v3-runtime commerce-1-v3-serve \
  --model-dir "$COMMERCE_ONE_MODEL_DIR" \
  --base-dir "$COMMERCE_ONE_MODEL_DIR" \
  --offline \
  --host 127.0.0.1 \
  --port 8000

The native request/response schema and a commerce example are in USAGE.md. The scorer is exposed at POST /v1/systemone and POST /api/alpha/decisions; a strict JSON-only bridge is exposed at POST /v1/chat/completions.

Measured evidence

Evaluation	Result	Boundary
Decision Index 0.2.1 complete union	60.99	Self-hosted reproduction of the complete public suite
Completed rows	150,317 / 150,317	Unanswered rows are not removed
Official Decision Index 0.3 Full Score	Pending maintainer evaluation	Not yet an official score or rank

The initial release is bound to the complete public 0.2.1 reproduction and an exact offline native-runtime qualification. Maintainers require immutable public model, runtime, and result links before they can run private 0.3 tests, so the Full Score remains pending at initial publication.

Qualified runtime

Field	Measured value
Hardware	1× NVIDIA H200
Precision	bfloat16
Runtime	PyTorch 2.14.0+cu130, CUDA 13.0
Input limit	65,536 tokens
Question microbatch	1
Fixture coverage	3 fixtures: Choice, Noul, Score
Native/direct parity	Exact; maximum absolute difference 0.0
Release/evaluated parity	Exact; maximum absolute difference 0.0
Throughput	8.836 decisions/s
Latency	340.146 ms p50, 349.026 ms p95 per complete fixture set
Network during qualification	Disabled

These figures describe the exact locked H200 qualification fixture and are not general serving benchmarks. Other accelerators, batching strategies, traffic shapes, and prompt lengths can produce different results. The complete safe attestation is QUALIFICATION.json.

Architecture

• Immediate parent: perplexity-ai/pplx-decider-v1.1-27b, immutable revision 3b45dead91dfa6d95aad6b95764a606fab2bf7a6.
• Ultimate base family: Qwen/Qwen3.8-27B, immutable revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.
• Native sharded Transformers checkpoint with a separate decision readout.
• Noncausal full-attention decision backbone with last-position pooling.
• Per-primitive held-out temperature calibration for Choice, Noul, and Score.
• Exact 65,536-token input limit with fail-closed overflow handling.
• Content-bound offline runtime; local bytes are verified before source import or model load.

Package contents

Path	Purpose
model/	Sharded BF16 checkpoint, tokenizer, and decision configuration
calibration/	Bound held-out temperatures for the three primitives
public-runtime/	Content-verified serving source closure
source/autojev/	Minimal decision-model source closure required by the checkpoint
evidence/	Sanitized evaluation and lineage attestations
training-data/lineage.json	Content-free training-source identity receipt
runtime.json	Locked runtime and backend identity
release-manifest.json	Exact package inventory and identities
QUALIFICATION.json	Content-free native runtime qualification
SHA256SUMS	Package checksum inventory

See ARCHITECTURE.md, LOCAL_RUN.md, API.md, and REPRODUCIBILITY.md.

Training data and provenance

The exact model lineage, required training-source identities, and source notice requirements are in PROVENANCE.json. Upstream notices are retained in THIRD_PARTY.md. These documents identify the release inputs without publishing protected training rows or private control receipts.

Release identity

Field	Verified value
Evaluated checkpoint SHA-256	b7fba0cceb1b57834fecacfbcb5e361d94fa185e877e891f4e43fc25c6458a1d
Release checkpoint SHA-256	f97262c59b102a39ae37826ffe1e1db8ddabaa1ec95b4e73bcb44328420ace57
Checkpoint transform attestation	ac7ae1330b4e912697b11ddb41770121f37ff88047e5f2046e3a338f9307d561
Native runtime identity	8a2fa16c5419da709a3cbe26ea2339f7ad4604db16da76f9c5426bd5a6984739
Public evaluation attestation	0e9432adcde227639ae7427bdcf9b3fdd3fb6b023d844d118a3cf23c3aa44749
Training-data lineage	d7a3c00a5fc71c30d10ea7cf92943ddc8a62925a276680aa9a862be53f970827
Public runtime qualification	2b01ab21b12c1d1c555688f205717708ade4a1b0a5b415aec08d36148c8e3639
Publication gate	INITIAL_PUBLIC_RELEASE_PENDING_OFFICIAL_V0_3

The release checkpoint differs from the evaluated checkpoint only by the documented safe metadata projection; qualification verifies identical decision outputs. No generic hosted-inference widget is enabled because Commerce-1 uses a typed custom runtime rather than free-form text generation.

License and contact

The project license is in LICENSE. Review LICENSE, PROVENANCE.json, and THIRD_PARTY.md before redistribution or deployment. Report runtime or model-card issues through the InferCrane Commerce-1 issue tracker.