Perplexity 发布 Decider v1.3:开放权重决策模型新增 Locate
Perplexity 发布了 pplx-decider-v1.3 和 pplx-decider-flash-v1.3,这是其开放权重决策模型的最新版本。Perplexity Decider v1.3 是一个决策模型,返回的是概率,而非生成的文本。它针对文本或图像输入回答是/否题、选择题和评分题。此次最主要的新增功能是 Locate,可按类型或指令提取精确的文本片段。权重以 Apache 2.0 协议在 Hugging Face 上发布。这两个模型也可在 Perplexity Decisions API 和 OpenRouter 上运行。你可以通过我们的 GPULeaderboard.com Chrome 扩展查看部署方案和规格。
Decider 是 Perplexity 的决策模型系列,灵感来自 TypeSafe 的 Jev。每个请求都包含一个共享的“状态(state)”,以及一个或多个带标注选项的问题。模型返回每个选项的概率,而不是写出答案。v1.3 接替 v1.2 和 Flash v1.2。Flash v1.3 是从大模型蒸馏而来。
Perplexity 将一个生成式 LLM 主干改造为分类器。输出层从基础模型的 LM head 出发,剪枝到选项 Token(A、B、C)。一个线性决策头读取 <|marker|> Token 处的隐藏状态,并输出各选项的概率。
全注意力层被改为双向,Perplexity 的消融实验表明这样做是有帮助的。
v1.3 还改变了问题之间的注意力机制。在 v1.2 中,每个问题都能看到其他所有问题,因此答案可能发生偏移。在 v1.3 中,状态能看到所有问题,但每个问题只能看到状态和自身。针对同一输入的所有问题仍在单次前向传播中运行。
训练分 3 个阶段进行:
权重在 3 个随机种子上取平均。Flash 通过最小化与大模型输出之间的 KL 散度来训练。
Locate 返回带起始和结束字符偏移量的精确文本片段。字段既可以是简短标签(“location”),也可以是指令(“Extract the city name where the job is located.”)。它面向命名实体识别(NER)、个人身份信息(PII)检测和抽取式问答。
每请求一个字段,就会新增一个 <|locate|> token。训练时只训练 span 头加一个 LoRA 适配器,主干网络和决策头保持冻结。适配器不做合并,因为合并会损害决策准确率。
下文所有结果均由 Perplexity 自行报告:
基准测试各行数据是 Perplexity 在其发布文章中自行报告的数字。
API 端点为 POST https://api.perplexity.ai/v1/decisions。模型 ID 为 pplx-decider-v1.3 和 pplx-decider-flash-v1.3,并提供 -latest 别名。定价为每 1M 输入 Token 0.02 美元和 0.01 美元。输出免费。
Perplexity 的延迟测试显示,一个约 243K Token 的请求在 v1.3 上耗时 16 秒,在 Flash 上耗时 5 秒。Locate 请求最多允许 128 个问题。
在本地,可通过 pplx_decider 包中的 Decider.from_pretrained() 加载模型。全注意力层必须以非因果(noncausally)方式运行。Perplexity 建议对细致的评分标准使用 v1.3,对高吞吐路由使用 Flash。
可查看技术细节、Decider v1.3 模型卡、Flash v1.3 模型卡和 API 文档。你也可以通过我们的 GPULeaderboard.com Chrome 扩展在此查看部署选项和规格。
需要与我们合作推广您的 GitHub 仓库、Hugging Face 页面、产品发布或网络研讨会等吗?请联系我们
文章《Perplexity 发布 Decider v1.3:开放权重决策模型新增 Locate》首先出现在 MarkTechPost 上。
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Perplexity Releases Decider v1.3: Open-Weight Decision Models Add Locate)
Perplexity has released pplx-decider-v1.3 and pplx-decider-flash-v1.3, the latest versions of its open-weight decision models. Perplexity Decider v1.3 is a decision model that returns probabilities, not generated text. It answers yes/no, multiple-choice and score questions about a text or image input. The main addition is Locate, which extracts exact text spans by type or instruction. Weights ship under Apache 2.0 on Hugging Face. Both models also run in the Perplexity Decisions API and on OpenRouter. You can check the deployment options and specs here through our GPULeaderboard.com chrome extension.
TL;DR
- Size: 27B on Qwen3.8-27B; Flash is 4B on Qwen3.5-4B. Hugging Face lists 26B and 5B params. API input: under 256K tokens. Local default: 32,768.
- Runs on: a CUDA GPU fitting ~49 GiB of BF16 weights (v1.3) or ~8.5 GiB (Flash). No quantized variants.
- Performance: leads TypeSafe’s Jev 1.14 on most of Perplexity’s self-reported evaluations.
- Best: 91.6% mean accuracy on 63 query-classification tasks.
- Worst: Locate averages 69.0, its lowest absolute score. Flash trails Jev 1.14 in finance, programming and law.
- Bottom line: Best: a 0.27% answer-flip rate and $26.47 per 1M production requests. Worst: every figure is self-reported.
What is Perplexity Decider v1.3?
Decider is Perplexity’s family of decision models, inspired by TypeSafe’s Jev. Each request has a shared ‘state’ plus one or more questions with labeled options. The model returns a probability per option instead of writing an answer. v1.3 succeeds v1.2 and Flash v1.2. Flash v1.3 is distilled from the large model.
How does Decider v1.3 work?
Perplexity adapts a generative LLM backbone into a classifier. The output layer starts from the base model’s LM head, pruned to option tokens (A, B, C). A linear decision head reads the hidden state at a <|marker|> token and outputs option probabilities.
Full-attention layers are made bidirectional, which Perplexity’s ablations showed helps.
v1.3 also changes attention between questions. In v1.2, every question saw every other question, so answers could shift. In v1.3, the state sees all questions, but each question sees only the state and itself. All questions about one input still run in a single forward pass.
Training runs in 3 stages:
- a full fine-tune on hundreds of public datasets
- a LoRA specialization adapter
- a LoRA vision adapter
Weights are averaged across 3 seeds. Flash is trained by minimizing KL divergence to the large model’s outputs.
What does the new Locate feature do?
Locate returns exact text spans with start and end character offsets. Fields can be short labels (‘location’) or instructions (‘Extract the city name where the job is located.’). It targets NER, PII detection and extractive QA.
A new <|locate|> token is added per requested field. A span head plus a LoRA adapter is trained, while the backbone and decision head stay frozen. The adapter stays unmerged because merging hurt decision accuracy.
How does Decider v1.3 perform on benchmarks?
All results below are self-reported by Perplexity:
- Query classification (63 tasks): 91.6% mean accuracy, 6.54 points above production classifiers.
- Domain classification: leads Jev 1.14 in all 7 domains, by up to 12.9 points in finance.
- Retrieval and reranking (72 tasks): 83.99 mean nDCG@10. That is 1.81 above Qwen3-Reranker-8B and 2.24 above Jev 1.14.
- Multilingual (19 tasks, 8 languages): 88.16% for v1.3, 86.50% for Flash, 85.41% for Jev 1.14.
- Locate (17 out-of-domain tasks): 69.0 average for v1.3 and 63.6 for Flash. Liquid d1 scores 43.3 and GLiNER large v2.5 scores 33.6.
- Image classification (5 tasks): 76.99% for v1.3, 1.47 points above Jev 1.14. Flash scores 72.77%.
- Real-world requests (1M requests, 2,318,887 questions): 93.98% agreement with frontier LLM judges. Jev 1.14 scores 93.35%.
- Robustness: answer flips on 0.27% of reorder or relabel transformations. Jev 1.14 flips on 2.69% to 4.84%.
How does Decider v1.3 compare with Jev 1.14 and Qwen3-Reranker-8B?
| Feature | Decider v1.3 | Decider Flash v1.3 | Jev 1.14 | Qwen3-Reranker-8B |
|---|---|---|---|---|
| Size | 27B per docs (HF lists 26B) | 4B per docs (HF lists 5B) | Not disclosed | 8B |
| Context | API under 256K; local default 32,768 | API under 256K; local default 32,768 | 262,144 | 32K |
| License | Apache 2.0 | Apache 2.0 | Not disclosed (hosted API) | Apache 2.0 |
| Modality | Text, images, Locate (text only) | Text, images, Locate (text only) | Text, images | Text |
| Hardware | ~49 GiB BF16 weights + working memory | ~8.5 GiB BF16 weights + working memory | Hosted only | Not disclosed |
| Price | $0.02 / 1M input, output free | $0.01 / 1M input, output free | $0.042 / 1M input, output free | Open weights, self-host |
| Retrieval nDCG@10 (72 tasks) | 83.99 | Not disclosed | 81.75 (derived from reported 2.24 gap) | 82.18 (derived from reported 1.81 gap) |
| Multilingual accuracy (19 tasks) | 88.16% | 86.50% | 85.41% | Not reported |
| Judge agreement (1M requests) | 93.98% | Not disclosed | 93.35% | Not reported |
| Cost for 1M requests | $26.47 | $13.24 | $68.10 | Not reported |
Benchmark rows are Perplexity’s self-reported numbers from its launch post.
How can developers deploy Decider v1.3?
The API endpoint is POST https://api.perplexity.ai/v1/decisions. Model IDs are pplx-decider-v1.3 and pplx-decider-flash-v1.3, with -latest aliases. Pricing is $0.02 and $0.01 per 1M input tokens. Output is free.
Perplexity’s latency test put an ~243K-token request at 16s on v1.3 and 5s on Flash. Locate requests allow up to 128 questions.
Locally, load the model with Decider.from_pretrained() from the pplx_decider package. Full-attention layers must run noncausally. Perplexity recommends v1.3 for nuanced rubrics and Flash for high-volume routing.
Key Takeaways
- Decider v1.3 returns probabilities, not text, for classification and scoring.
- New Locate extracts exact spans for NER, PII and extractive QA.
- v1.3 scores 91.6% on 63 query-classification tasks (self-reported).
- API pricing is $0.02 (v1.3) and $0.01 (Flash) per 1M input tokens.
- Weights are Apache 2.0; v1.3 needs ~49 GiB, Flash ~8.5 GiB.
Check out the technical details, Decider v1.3 model card, Flash v1.3 model card and API docs. You can check the deployment options and specs here through our GPULeaderboard.com chrome extension.
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.
Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us
The post Perplexity Releases Decider v1.3: Open-Weight Decision Models Add Locate appeared first on MarkTechPost.
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 MarkTechPost(Perplexity Releases Decider v1.3: Open-Weight Decision Models Add Locate),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。