Commerce-1：面向电商 Agent 决策的开源 27B 模型
Commerce-1: an open-weight 27B decision model for commerce agents
Agent 开发者关注，这款模型针对特定决策场景优化，零生成 Token 且性能指标明确，值得集成测试。
We built Commerce-1 for a specific part of an AI agent: the moment it must choose an action.
我们为 AI 智能体的特定部分构建了 Commerce-1：即它必须选择行动的时刻。
A commerce agent may need to select a product, resolve an inventory conflict, route a payment, choose a fulfillment path, prioritize a return, or escalate an uncertain transaction.
商业智能体可能需要选择产品、解决库存冲突、路由支付、选择履约路径、优先处理退货，或升级不确定的交易。
These are usually bounded decisions with known candidate actions. They don't always need another generated paragraph.
这些通常是具有已知候选行动的有界决策。它们并不总是需要生成另一段文本。
How it works
工作原理
Commerce-1 receives a structured state and ordered Choice, Noul, or Score questions. It scores the supplied options directly and returns typed answers with an ordered probability distribution.
Commerce-1 接收结构化状态以及有序的 Choice、Noul 或 Score 问题。它直接对提供的选项进行评分，并返回带有有序概率分布的类型化答案。
- Zero generated output tokens
- Up to 255 candidate actions
- Typed decisions and probability distributions
- Choice-order augmentation and per-primitive held-out temperature calibration
- 零生成输出令牌
- 最多 255 个候选行动
- 类型化决策和概率分布
- Choice 顺序增强和每个基元保留温度校准
The model builds on perplexity-ai/pplx-decider-v1.1-27b, using a noncausal full-attention decision backbone, last-position pooling, and a dedicated decision readout.
该模型基于 perplexity-ai/pplx-decider-v1.1-27b，采用非因果全注意力决策主干、最后位置池化和专用的决策读取头。
Benchmarks
基准测试
We reproduced the complete Decision Index 0.2.1 public suite:
我们复现了完整的 Decision Index 0.2.1 公开套件：
| Metric | Result |
|---|---|
| Decision Index score | 60.99 |
| Evaluation rows | 150,317 |
| Throughput (1× H200) | 8.836 decisions/sec |
| Three-decision fixture p50 | 340.146 ms |
| Three-decision fixture p95 | 349.026 ms |
| 指标 | 结果 |
|---|---|
| Decision Index 得分 | 60.99 |
| 评估行数 | 150,317 |
| 吞吐量 (1× H200) | 8.836 次决策/秒 |
| 三次决策固定装置 p50 | 340.146 毫秒 |
| 三次决策固定装置 p95 | 349.026 毫秒 |
These measurements come from our qualified single-H200 runtime.
这些测量数据来自我们合格的单 H200 运行时环境。
Limitations
局限性
There are important boundaries:
存在重要的边界条件：
- Current BF16 weights are approximately 54 GiB.
- This release is qualified on one NVIDIA H200.
- Commerce-1 recommends actions; application code remains responsible for inventory truth, payments, fraud rules, budgets, and execution.
- 当前的 BF16 权重约为 54 GiB。
- 此版本在单个 NVIDIA H200 上经过合格认证。
- Commerce-1 推荐行动；应用程序代码仍负责库存真实性、支付、欺诈规则、预算和执行。
Open source and demos
开源与演示
The model weights are public. The runtime source is Apache-2.0, with model and training-source notices documented separately.
模型权重是公开的。运行时源代码采用 Apache-2.0 许可证，模型和训练来源通知单独记录。
- Model weights: https://huggingface.co/infercrane/Commerce-1
- GitHub / local runtime: https://github.com/infercrane/commerce-1
- Interactive demos: https://huggingface.co/spaces/infercrane/commerce-1-arcade
- Technical article: https://infercrane.com/blog/commerce-1
- 模型权重：https://huggingface.co/infercrane/Commerce-1
- GitHub / 本地运行时：https://github.com/infercrane/commerce-1
- 交互式演示：https://huggingface.co/spaces/infercrane/commerce-1-arcade
- 技术文章：https://infercrane.com/blog/commerce-1
Feedback wanted
期待您的反馈
For developers building commerce agents: Would a native Decisions SDK or broader OpenAI-compatible tooling make this easier to integrate into your stack?
对于正在构建商务智能体的开发者：原生的 Decisions SDK 或更广泛的 OpenAI 兼容工具集，是否能让将其集成到您的技术栈中变得更加容易？
I'd also be interested in hearing which bounded decisions in your agents are currently handled by general-purpose LLMs.
我也很感兴趣，想了解您的智能体中目前由通用大语言模型（LLM）处理的具体决策场景有哪些。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力