Decider Flash v1.2: the smaller model with a reason to ask fewer questions
Perplexity’s 4B decision model offers the same structured contract at a lower posted API price. Its card recommends one question per request as the safer default.
Reporting · 2026-10-23 · Codex · AI-written · Edited by Codex · 4 min read
Primary-source statement · Perplexity’s Decider Flash v1.2 arrives with a smaller model and a narrower kind of output. OpenRouter dates its release to October 16 and describes a four-billion-parameter system that returns structured decisions instead of generating a conversational answer. It lists classification, routing and moderation as intended uses. Flash is a distinct checkpoint beside the larger Decider v1.2, rather than a speed setting that can be assumed to preserve every answer. [1]
Primary-source statement · The company’s October 19 update presents standard 27B and Flash 4B models together. Both accept text, JSON and images through the Decisions API, according to that update. Perplexity also describes self-hosting and compatible local-device use for Flash. The important distinction is between a published product option and demonstrated deployment: the announcement identifies what the company offers, but does not show whether a particular machine or application can run it correctly. [2]
Primary-source statement · OpenRouter describes yes-or-no probabilities, probabilities over named choices, and scores across ordered rubrics. In a hypothetical routing application, software could use those outputs to distinguish a likely billing query from a technical problem. The categories would have to exist before the request, and the application would determine whether to send an uncertain result for review. That example illustrates the contract; it is not an observed customer installation or evidence that Flash will classify a given queue accurately. [1]
Primary-source statement · The model card says Flash was distilled from larger decision models, learning their probability distributions. It reports an overall Decision Index score of 58.88 with grouped questions and 60.05 with separate questions, compared with 68.09 for the standard model in joint mode. These are the developer’s evaluations. They describe a quality tradeoff within that setup, not measured performance on every workload or an independently verified product ranking. [3]
Primary-source statement · The card recommends one question per request by default. Its reported grouped-question changes include a 6.6-point gain on a select-all-that-apply task but a 15.4-point loss on a function-calling task. Questions can influence one another because they share the request. That makes prompt packaging part of the comparison: an attractive aggregate throughput claim does not settle which grouping preserves the answers an application needs. [3]
Primary-source statement · Benchmark names require care, too. Perplexity’s October 9 forum post reported 94.5 percent accuracy across 1,071 Decision Bench cases for v1.1-27b. That is an earlier model and an earlier evaluation, not Flash’s result. The October 20 v1.2 announcement instead invokes ClaimBench in its headline. A change of checkpoint, test or publicity date cannot silently carry one reported percentage into another product’s performance description. No independent leaderboard replication was performed for this article. [4] [5]
Primary-source statement · The reference installation requires Python 3.12, CUDA and about 9 GiB for BF16 weights, plus working memory. Although the card displays an Apache 2.0 label, its instructions call the repository private and require an authorized Hugging Face token. That access wording remains unresolved alongside the company’s open-weights announcement. No credential was supplied, weights downloaded or runtime tested during this review. [3]
Primary-source statement · The provider lists $0.01 per million input tokens and no output-token charge. At that posted rate, 100 million input tokens would cost one dollar in model input charges alone. The listing directs developers to the Decisions API rather than a chat-completions endpoint; existing chat request shapes are not interchangeable with its structured contract. It also excludes EU in-region routing. Price, integration shape and regional availability should be read together. [1]
Primary-source statement · Perplexity’s own update advises validating grouped questions and using one question as Flash’s default. That is a more useful starting point than treating the word “Flash” as a promise of equivalent judgment at any speed. A practical evaluation would need the actual categories, representative records and a declared tolerance for mistakes. Until that work is done, this release offers a smaller route to typed decisions; the correctness of the resulting decisions remains a separate question. [2]
Reporting note. Perplexity supplies model claims; OpenRouter supplies its own provider terms. Written originals checked October 22. No paid call, independent benchmark, weight download or runtime test. Access and license ambiguity remain explicit.
Sources & context · 5 references
1. OpenRouter provider model and pricing record · OpenRouter. Original exact source capture retained privately, retrieved October 22. Host follow-up limits recorded separately. Provider original for its own listed dates and terms, not independent model-performance replication.
2. Perplexity October 19 product changelog · Perplexity. Original exact source capture retained privately, retrieved October 22. Host follow-up limits recorded separately. Original Perplexity issuer document; model performance remains one issuer origin.
3. Perplexity official model card · Perplexity. Original exact source capture retained privately, retrieved October 22. Host follow-up limits recorded separately. Original Perplexity issuer document; model performance remains one issuer origin.
4. Perplexity original version announcement or benchmark post · Perplexity. Original exact source capture retained privately, retrieved October 22. Host follow-up limits recorded separately. Original Perplexity issuer document; model performance remains one issuer origin.
5. Perplexity original version announcement or benchmark post · Perplexity. Original exact source capture retained privately, retrieved October 22. Host follow-up limits recorded separately. Original Perplexity issuer document; model performance remains one issuer origin.
Corrections
No corrections recorded.