Perplexity open-sources two decision models: per million inputs Token as low as 1 cents, with inference speed up to 5 times
BIBIBI
AT A GLANCE
Perplexity open-sources two multimodal decision models, with input prices as low as per million Token 0.01 USD。
Article
Perplexity open-sources two decision models: per million inputs Token as low as 1 cents, with inference speed up to 5 times
Beating AI news flash, Perplexity open-sources two multimodal decision models pplx-decider-v1.2（27B) and pplx-decider-flash-v1.2（4B), based respectively on Qwen3.8-27B and Qwen3.5-4B. They are not designed to generate long-form answers; instead, they classify, score, and make judgments based on text or images, and can be used for Agent task routing, content moderation, and other scenarios.
Compared with the previous generation's question-by-question inference, the new version allows multiple questions to share the same input and completes the judgments in a single inference pass. Perplexity says inference speed improves by up to 5 times in some scenarios. However, 4B version may suffer a loss in accuracy when combining multiple questions, and the official recommendation is to process questions individually by default.
In the public tests of the Decision Index 0.3, the 27B model's score rose from 62.30 in the previous generation to 68.09, while the 4B version ranked first among models of the same size. However, the new version's training included official training sets from some evaluation tasks.
Both models use the Apache 2.0 license and are available through Perplexity API and Vercel AI Gateway。API at a per-million-input price of Token respectively, while output is free.
Original link 0.02 USD and 0.01 USD https://m.theblockbeats.info/flash/372319
Key points
01
Perplexity has open-sourced pplx-decider-v1.2（27B) and pplx-decider-flash-v1.2（4B) two multimodal decision models.
02
The two models are based respectively on Qwen3.8-27B and Qwen3.5-4B。
03
The models classify, score, and make judgments based on text or images, and can be used for Agent task routing, content moderation, and other scenarios.
04
The new version supports multiple questions sharing the same input and completing the judgments in a single inference pass.
05
Perplexity says inference speed improves by up to 5 times in some scenarios.
06
The report notes that the 4B version may suffer a loss in accuracy when combining multiple questions, and the official recommendation is to process questions individually by default.
07
In the Decision Index 0.3 public tests, the 27B model's score rose from 62.30 to 68.09，4B version ranked first among models of the same size.
08
The new version's training included official training sets from some evaluation tasks.
09
The two models use the Apache 2.0 license and are available through Perplexity API and Vercel AI Gateway。
10
API per million inputs Token are priced at 0.02 USD and 0.01 USD respectively, while output is free.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
These two models are more like “judges”: their focus is not writing long answers, but classifying, scoring, and making judgments about text or images. The new version can also let multiple questions reuse the same input to reduce redundant inference; however, 4B version may sacrifice accuracy when handled this way.
Why it matters to readers
Open-source licenses, low input prices, and shared-input capabilities may lower the barrier to using these models for classification, content moderation, and Agent task routing. However, whether the speed and leaderboard results can be replicated in real-world business operations still requires further validation.
Beginners can think of it as a specialized AI model for routing and moderation. When choosing, do not look only at price and leaderboard rankings; also compare accuracy and latency on real tasks, as well as 4B model's change in precision when making batch judgments.
Risks and unknowns
Highest 5 The claim of a times speedup is Perplexity's wording; evidence of independent test results was not provided.
The training data incorporated official training sets from some evaluation tasks, which may affect the comparability of public test results.
The evidence did not provide model weights, complete evaluation details, or an official announcement link.
Accuracy, latency, concurrency capability, and cost performance in real-world business scenarios have not been specified.
Related Developments
Loading event timeline…
Apache 2.0 The scope of the license and the actual availability of the two access channels have not yet been verified.
Related concepts
Token
This term is not in the glossary yet. Browse related concepts in the glossary.
