Supersonic Labs发布1.44亿参数开源决策模型Julia 1
Supersonic Labs Releases Julia 1: A 144.3M-Parameter Open Decision Model That Runs on a CPU
极低成本的开源决策模型落地实践,适合需要端侧部署或低延迟推理的Agent路由场景,建议开发者关注其架构与性能权衡。
Supersonic Labs, a small AI lab from Brazil, has released Julia 1. It is a compact decision model, not a chatbot. You pass it context, a question, and 2 to 20 candidate answers. It picks one and returns a probability for every option. The model has 144.3M parameters and runs on a plain CPU.
来自巴西的小型 AI 实验室 Supersonic Labs 发布了 Julia 1。它是一个紧凑的决策模型,而非聊天机器人。你向其输入上下文、问题以及 2 到 20 个候选答案。它会选择一个答案,并为每个选项返回一个概率。该模型拥有 1.443 亿参数,可在普通 CPU 上运行。
Is it deployable? Yes. The weights are on Hugging Face under Apache 2.0 and run locally with Python 3.11+ on CPU or a BF16-capable GPU. An ONNX build also runs in the browser via WebGPU. A hosted API is announced but not open yet.
可以部署吗?可以。权重以 Apache 2.0 许可证托管在 Hugging Face 上,可在配备 Python 3.11+ 的 CPU 或支持 BF16 的 GPU 上本地运行。ONNX 版本也可通过 WebGPU 在浏览器中运行。已宣布提供托管 API,但尚未开放。
What Julia 1 Does
Julia 1 的功能
Julia 1 handles three decision types through one API:
Julia 1 通过单一 API 处理三种决策类型:
- choice: pick one label from 2 to 20 described options (classification, routing).
- score: return the expected index on an ordered rubric, such as low, medium, high.
- noul: return the probability that a yes-or-no statement is true.
- choice:从 2 到 20 个描述性选项中选择一个标签(分类、路由)。
- score:在有序评分标准(如低、中、高)上返回预期索引。
- noul:返回是非陈述为真的概率。
Results come back in the caller’s option order with full softmax probabilities. Caller IDs such as billing are returned unchanged. The model does not generate text.
结果按调用方的选项顺序返回,并附带完整的 softmax 概率。调用方 ID(如 billing)原样返回。该模型不生成文本。
Architecture and Training Budget
架构与训练预算
Julia 1 starts from JHU CLSP’s mmBERT-small, a 140M-parameter multilingual ModernBERT encoder trained on 1,800+ languages. Supersonic Labs kept the encoder and tokenizer, added a decision head, and trained on decision-format examples. The lab states Julia 1 is not a fine-tuned Qwen model. The runtime supports 8,192 combined tokens, but published benchmarks used a 1,024-token limit.
Julia 1 基于 JHU CLSP 的 mmBERT-small,这是一个在 1,800 多种语言上训练的、拥有 1.4 亿参数的多语言 ModernBERT 编码器。Supersonic Labs 保留了编码器和分词器,添加了决策头,并在决策格式示例上进行了训练。该实验室声明 Julia 1 并非 Qwen 模型的微调版本。运行时支持 8,192 个组合令牌,但发布的基准测试使用了 1,024 个令牌的限制。
Total cloud GPU spend for training and experiments was about R$540 (US$104.08). The FP32 weights occupy 550.5 MiB. The private training pipeline is not released. Julia 2, with the lab’s own foundation architecture, is in development.
训练和实验的总云 GPU 支出约为 540 雷亚尔(104.08 美元)。FP32 权重占用 550.5 MiB。私有训练管道未发布。Julia 2 正在开发中,将采用该实验室自有的基础架构。
Benchmark Results
基准测试结果
The September 24, 2026 evaluation ran on H200 BF16 with strict encoding. The comparison baseline is TypeSafe’s Jev, using reference values from the Jev benchmark protocol, not a new Jev run.
2026 年 9 月 24 日的评估在 H200 BF16 上进行,采用严格编码。对比基线是 TypeSafe 的 Jev,使用来自 Jev 基准协议参考值,而非新的 Jev 运行结果。
- Typed Decisions: 73.15% (1,463/2,000) vs 72.70% reference.
- AG News, 4 labels: 94/100 vs 91% reference.
- DAIR Emotion, 6 labels: 86/100 vs 48% reference.
- Banking77, 72 labels: 64/100 vs 87% reference. This is the clear failure.
- MASSIVE, 18 scenarios: 71.50% macro accuracy across 52 locales; 86.25% pt-PT, 86.75% en-US.
- Typed Decisions:73.15%(1,463/2,000),对比参考值 72.70%。
- AG News,4 个标签:94/100,对比参考值 91%。
- DAIR Emotion,6 个标签:86/100,对比参考值 48%。
- Banking77,72 个标签:64/100,对比参考值 87%。这是明显的失败案例。
- MASSIVE,18 个场景:跨 52 个地区的宏观准确率为 71.50%;pt-PT 为 86.25%,en-US 为 86.75%。
The classification pilots use only 100 examples each. A September 25 CPU run reproduced most numbers: 72.55% on Typed Decisions and 60/100 on Banking77 with 3 abstentions.
分类试点仅各使用 100 个示例。9 月 25 日的 CPU 运行复现了大部分数值:Typed Decisions 为 72.55%,Banking77 为 60/100(含 3 次弃权)。
On-Device Latency
设备端延迟
The lab published per-device measurements. On an Apple M4, one decision per call took a 33.15 ms median. On a Samsung SM-X510 tablet via ONNX Runtime, the median was 203 ms with 393.1 MB peak RSS. On an Intel Core i5-1235U, AG News decisions took a 107.83 ms median. Banking77 took 3,713.54 ms because it narrows 72 labels first.
实验室发布了每台设备的测量数据。在 Apple M4 上,每次调用做出一次决策的中位时间为 33.15 毫秒。通过 ONNX Runtime 在三星 SM-X510 平板上,中位时间为 203 毫秒,峰值 RSS 为 393.1 MB。在 Intel Core i5-1235U 上,AG News 决策的中位时间为 107.83 毫秒。Banking77 耗时 3,713.54 毫秒,因为它首先将标签范围缩小至 72 个。
On X, @supersonicai claims Julia 1 classifies 5x faster than Jev on an i5 laptop. Treat that carefully. The Jev pilot measured Jev as a hosted service called from France, so latencies are not like-for-like.
在 X(原 Twitter)上,@supersonicai 声称 Julia 1 在 i5 笔记本电脑上的分类速度比 Jev 快 5 倍。请谨慎对待此说法。Jev 的试点项目将 Jev 作为从法国调用的托管服务进行测量,因此延迟情况并非完全对等。
Introducing Julia-1:
介绍 Julia-1:
Our first classification model that runs on almost anything.
我们的首款几乎能在任何设备上运行的分类模型。
Learn more https://t.co/YJCdEeIBSo pic.twitter.com/3cizsbG9ZB
了解更多 https://t.co/YJCdEeIBSo pic.twitter.com/3cizsbG9ZB
— Supersonic Labs (@supersonicai) September 26, 2026
—— Supersonic Labs (@supersonicai) 2026年9月26日
Interactive Explainer
交互式解释器
Julia 1 vs Closest Competitors
Julia 1 与最接近的竞争对手对比
| Feature | Julia 1 | TypeSafe Jev | GLiNER2.5 Multi |
|---|---|---|---|
| Developer | Supersonic Labs | TypeSafe AI | Fastino |
| Access | Open weights | Hosted API, early access | Open weights |
| License | Apache 2.0 | Proprietary | Apache 2.0 |
| Parameters | 144.3M | Not disclosed | 287M |
| Base encoder | mmBERT-small | Not disclosed | mDeBERTa-v3-base |
| Decision types | Choice, score, yes/no | Typed structured decisions | Classification, NER, relations, records |
| Options per call | 2 to 20 (Router for more) | Up to 255 | Label list per schema |
| Runs locally on CPU | Yes | No | Yes |
| Input price per 1M tokens | $0.025 (planned API) | $0.042 | Free (self-hosted) |
| AG News pilot | 94% | 91% | 70% |
| DAIR Emotion pilot | 86% | 48% | 44% |
| Banking77 pilot | 64% | 87% | 61% |
| 特性 | Julia 1 | TypeSafe Jev | GLiNER2.5 Multi |
|---|---|---|---|
| 开发者 | Supersonic Labs | TypeSafe AI | Fastino |
| 访问方式 | 开放权重 | 托管 API,早期访问 | 开放权重 |
| 许可证 | Apache 2.0 | 专有 | Apache 2.0 |
| 参数量 | 1.443亿 | 未披露 | 2.87亿 |
| 基础编码器 | mmBERT-small | 未披露 | mDeBERTa-v3-base |
| 决策类型 | 选择、评分、是/否 | 结构化类型决策 | 分类、命名实体识别、关系、记录 |
| 每次调用的选项数 | 2 到 20(更多需使用路由器) | 最多 255 | 每个模式的标签列表 |
| 可在 CPU 本地运行 | 是 | 否 | 是 |
| 每百万 token 输入价格 | $0.025(计划中的 API) | $0.042 | 免费(自托管) |
| AG News 试点结果 | 94% | 91% | 70% |
| DAIR Emotion 试点结果 | 86% | 48% | 44% |
| Banking77 试点结果 | 64% | 87% | 61% |
Sources: Julia 1 model card, TypeSafe launch post, GLiNER2.5 Multi card, Jev benchmark pilot. Julia 1 pilots ran separately from the Jev and GLiNER runs.
来源:Julia 1 模型卡片、TypeSafe 发布文章、GLiNER2.5 Multi 模型卡片、Jev 基准测试试点。Julia 1 的试点项目独立于 Jev 和 GLiNER 的运行进行。
Limitations
局限性
Julia 1 compares the answers you supply. It cannot be counted on for missing facts, algebra, or multi-step calculation. The Router can drop the correct label during narrowing. It is not a drop-in Transformers pipeline, and no Hugging Face inference provider serves it. Supersonic Labs advises evaluating on your own questions and keeping humans in the loop for consequential decisions.
Julia 1 比较你提供的答案。它不能用于处理缺失事实、代数或多步计算。路由器在缩小范围过程中可能会丢弃正确的标签。它不是即插即用的 Transformers 管道,也没有 Hugging Face 推理提供商提供服务。Supersonic Labs 建议在你自己的问题上进行评估,并在重大决策中保留人工审核环节。
Key Takeaways
关键要点
- Julia 1 is a 144.3M-parameter, Apache 2.0 decision model that runs on CPU.
- One API covers choice, ordered score, and yes-or-no decisions over 2 to 20 options.
- It beat Jev references on 3 of 4 pilots but trailed badly on 72-label Banking77.
- Median latency hit 33.15 ms per decision on an Apple M4.
- Training cost about US$104 in cloud GPUs; a $0.025/MTok API is planned.
- Julia 1 是一款拥有 1.443 亿参数、采用 Apache 2.0 许可证、可在 CPU 上运行的决策模型。
- 一个 API 即可覆盖 2 到 20 个选项的选择、有序评分和是非决策。
- 它在 4 项试点中的 3 项击败了 Jev 引用,但在拥有 72 个标签的 Banking77 试点中表现明显落后。
- 在 Apple M4 上,每次决策的中位延迟达到 33.15 毫秒。
- 训练成本约为 104 美元(云 GPU);计划推出单价为 $0.025/百万 token 的 API。
Check out the Model Weights, ONNX/WebGPU build, and Technical details. All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
查看模型权重、ONNX/WebGPU 构建版本和技术细节。所有功劳归于该项目的研究者。此外,欢迎在 Twitter 上关注我们,别忘了加入拥有超过 15 万成员的 ML SubReddit,并订阅我们的新闻通讯。等等!你在 Telegram 上吗?现在你也可以加入我们的 Telegram 群组了。
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