Perplexity Releases Decider v1.3 and Flash v1.3 Weights Under Apache 2.0; Artificial Analysis Decision Index Lists v1.3 Above Jev 1.14, Flash Below
MarkTechPost reports Perplexity's open-weight Decider v1.3 and Flash v1.3 ship under Apache 2.0 on Qwen bases; Artificial Analysis's Decision Index scores v1.3 at 70.41, Jev 1.14 at 66.23 and Flash v1.3 at 64.07.
Overview
Perplexity has released two open-weight decision models, pplx-decider-v1.3 and pplx-decider-flash-v1.3, MarkTechPost reported on November 13. A decision model returns probabilities for options the caller supplies instead of generated text, and v1.3 adds a feature called Locate that extracts exact text spans, according to the same report. The Hugging Face entries for both models list the licence as Apache 2.0. Artificial Analysis, which says its evaluations are run independently, places v1.3 above TypeSafe’s Jev 1.14 on its Decision Index and Flash v1.3 below it. Everything below is as read on November 15, 2026.
What was released
MarkTechPost reports that the weights ship under Apache 2.0 on Hugging Face and that both models also run in Perplexity’s Decisions API and on OpenRouter. It gives API input prices of $0.02 per million tokens for v1.3 and $0.01 for Flash, with output free.
The Hugging Face API entry for v1.3 shows the repository was created at 08:38 UTC on 2026-11-09 and last modified at 09:32 UTC on 2026-11-10; the entry for Flash v1.3 shows the same two days. The v1.3 entry names Qwen/Qwen3.8-27B as its base model, and the Flash entry names Qwen/Qwen3.5-4B. The Hub’s entries for Qwen3.8-27B and Qwen3.5-4B both list apache-2.0.
On licensing, each Perplexity repository carries a LICENSE file that is the text of the Apache License, Version 2.0 (v1.3, Flash). Each NOTICE file names its Qwen base and revision and says the work is derived from it “under Apache License 2.0” (v1.3, Flash). Neither model card lists licences for the training data; the v1.3 card credits the tasksource collection for a large part of the gains, and MarkTechPost says the first training stage used hundreds of public datasets.
Access and running it
Both model cards say: “The repository is private: authenticate with a Hugging Face token that has access to it.” The Hub API entries list private as false and gated as false, and the Herald’s own anonymous requests for each repository’s resolve/main/config.json returned the file on November 15. None of the pages the Herald read explains the difference.
The v1.3 card asks for Python 3.12 or later and a CUDA GPU with room for about 49 GiB of BF16 weights; the Flash card says about 8.5 GiB. The v1.3 card also says “Default causal inference does not reproduce this checkpoint”, and its usage example imports pplx_decider from the source folder bundled in the repository. The card also states that fast=True and locate=False disable Locate.
The Hub’s safetensors metadata counts 26,085,330,160 parameters for v1.3 and 4,539,265,536 for Flash. MarkTechPost’s table gives 27B and 4B “per docs” and notes that Hugging Face lists 26B and 5B. The Qwen3.8-27B and Qwen3.5-4B entries count 27,781,427,952 and 4,659,865,088. No page the Herald read explains the gaps; the v1.3 card says the artifact has a separate readout rather than a full-vocabulary lm_head, and the Herald has not tested whether that accounts for them.
MarkTechPost says that in v1.3 each question sees only the shared state and itself, and that all questions about one input still run in a single forward pass. The v1.3 card describes two library modes: by default (mode="separate") each question is answered in its own sequence, and mode="isolated" encodes a request as one sequence, in which, the card says, 1.5-2.2% of top answers differed from separate on its samples. On the Herald’s reading, the single-pass description matches the isolated mode rather than the default. Neither page says which one Perplexity’s hosted API uses.
Perplexity’s figures
MarkTechPost states that the benchmark results it lists are “self-reported by Perplexity”. They include 91.6% mean accuracy on 63 query-classification tasks, and MarkTechPost’s summary says v1.3 leads Jev 1.14 on most of Perplexity’s self-reported evaluations while Flash trails Jev 1.14 in finance, programming and law. The v1.3 card calls its default mode “the computation all the scores above were measured with”, but, as the Herald read it, the card contains no benchmark scores.
What Artificial Analysis lists
The Artificial Analysis Decision Index page states: “All evaluations are conducted independently by Artificial Analysis.” Its About page calls the company “the independent benchmarking company for AI”. Version 2.1 of the index weights six evaluations: AA-Paperclip 30%, EvalSafe-O*NET-AA 15%, When2Call 10%, AA-Omniscience-Decision 15%, HLE-MultiChoice 15% and AA-Symmetry 15%. The page text the Herald read shows no date, and the per-model figures sit in the page’s embedded data rather than its visible text. The Herald’s October 25 Analysis of decision-model evaluation said it could not read those values.
In that data, the index scores are 70.41 for Decider v1.3, 69.77 for v1.2, 66.23 for Jev 1.14 and 64.07 for Flash v1.3. Counting the 22 decision-model entries in the data, the Herald finds v1.3 sixth, Jev 1.14 eleventh and Flash v1.3 thirteenth; the data also holds four language-model entries, which would put v1.3 eighth. The page’s charts are labelled “18 of 22 models” by default.
By component (Decision Scores, except the Symmetry Score on AA-Symmetry), v1.3 is ahead of Jev 1.14 on AA-Paperclip (68.74 against 57.53), When2Call (94.62 against 84.39) and AA-Symmetry (77.93 against 74.50), and behind on EvalSafe-ONET-AA (89.09 against 91.14), AA-Omniscience-Decision (58.93 against 60.14) and HLE-MultiChoice (42.89 against 44.47). Flash v1.3 is ahead of Jev 1.14 on When2Call (90.14 against 84.39) and AA-Symmetry (80.37 against 74.50) and behind on the other four. Artificial Analysis describes EvalSafe-ONET-AA as based on TypeSafe’s O*NET benchmark; MarkTechPost attributes Jev to TypeSafe. On the page’s weighted median time per task, the data lists about 289 ms for v1.3, 178 ms for Flash v1.3 and 92 ms for Jev 1.14.
What the sources do not establish
- The Artificial Analysis page text the Herald read shows no date or run notes, so the Herald cannot say when the entries were added, and rankings may change.
- Artificial Analysis’s evaluations differ from the ones MarkTechPost relays from Perplexity, so the two sets of figures do not test the same tasks.
- The Herald’s counts and ranks are its own, taken from the page’s embedded data on November 15, not from a table printed on the page.
- No page the Herald read explains the private-repository wording or establishes the cause of the parameter-count gaps.