ENPT

pplx-decider v1.2 (27B + Flash 4B)

Last updated:

Open96–96ms$0.02/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity6.3/10

    Apache-2.0 weights on Hugging Face, a priced hosted API ($0.02/M; Flash $0.01/M) with quickstart, reference, cookbooks and 10 req/s per org. Needs the bundled code (noncausal attention, per-question readout). The HF card still says the repo is private. No SLA.

  • Capability8.1/10

    Noul / Choice / Score with images and a joint request format. Full 65.55 on the official Decision Index 0.3.2 (#4; Jev 60.11). Temperature fitted, ECE 0.014 on the dev set but about 0.05 held-out. 96 ms median is Perplexity's own GB200 run, no p95. DI training splits are in the data.

  • Adoption18/100

    Default model of Perplexity's Decisions API (old v1/v1.1 names now route here). Launch posts: CEO 623 likes / 32k views, Perplexity Devs 277 likes / 76k views. HF: 129 + 224 downloads, 19 + 16 likes in 1-2 days, several community quants. No named production users.

Vendor claims

publicado (OK do Fabricio 17/10 BRT).

Read this first.

  1. One page, two models. pplx-decider-v1.2 (27B) and pplx-decider-flash-v1.2 (4B) share a card layout, an API and a request format. The 27B is the headline; Flash has its own numbers below. The older pplx-decider v1.1 page stays for history.
  2. Perplexity put the Decision Index training splits in. The card says several Decision Index benchmarks ship official training splits and that “we now include these splits (decontaminated against the test sets) to make comparisons on the board fair”. The board’s Full score also includes the maintainers’ private tests, so we use that one; the 68.09 on the card is the public suite, run by Perplexity.
  3. The card still says “the repository is private”. It is public and ungated (we read it on 17 Oct). Same wording error as v1.1.
  4. Old API names now point here. The docs say requests naming pplx-decider-v1.1-27b or pplx-decider-v1-27b “select pplx-decider-v1.2 and bill at its rate”.

What changed from v1.1

  • Joint request format. v1.1 answered one question per forward pass, read at the last token. v1.2 reads the state and every question in one sequence, with a readout at a marker token after each question’s options. Several questions about one state cost about as much as one. Full attention stays noncausal, so questions see each other.
  • Grouping can hurt. The card’s own table: grouping helps when questions are facets of one decision (SATA-Bench +11.3, ACOS +7.8, PhishNChips +5.1) and hurts when they are competing candidates (ToolRet −4.8, BRIGHT −2.9, BFCL −1.1). Perplexity recommends one candidate per request when in doubt.
  • New data and souping. A reworked training set and several runs combined by weight averaging. Flash (Qwen3.5-4B) is distilled from the 27B decision models on their training requests, on full probability distributions.

Specs

Attribute 27B pplx-decider-v1.2 Flash pplx-decider-flash-v1.2
Developer Perplexity Perplexity
Base Qwen3.8-27B (26.1B params in safetensors) Qwen3.5-4B (4.5B params stored)
License Apache-2.0 Apache-2.0
Weights ~52 GiB BF16 + a BF16 [255, 5120] readout ~9 GiB BF16
Types choice (1–255 options), noul, score (2–255 levels on the card; the API says 10) same
Inputs text, images up to 1.6 MP same
Context 262,144 tokens per sequence on the card; trained up to ~32k; 256K in the API docs same
Calibration temperature 1.382 in decision_config.json temperature 1.554
Serving bundled pplx_decider package (reference and fast=True path) or the hosted API same
Hosted API POST /v1/decisions, alias pplx-decider-latest, $0.02 per 1M input tokens, output free alias pplx-decider-flash-latest, $0.01 per 1M
Rate limit 10 requests per second per organization, shared by both models
Released HF repos created 15 Oct 2026; announced on X 16 Oct HF repo 15 Oct; announced 16 Oct

Official board: Decision Index 0.3.2

Data generated 16 Oct 2026 19:00 UTC (the board lists v1.2 and Flash as new entries; it shows no v1.1 row any more). Full = 0.2 public + 0.5 same-skills + 0.3 new-domains.

Model Full Rank Public Public rank Same skills New domains
pplx-decider v1.2 (27B) 65.55 #4 67.74 #3 64.30 55.78
pplx-decider Flash v1.2 (4B) 54.81 #47 60.06 #22 56.16 43.17
Jev 1.13 60.11 #18 57.96 #37

For scale: v1.1 was #1 at 62.75 on the 6 Oct board. The v1.2 jump on the public suite (62.30 → 68.09 on the card) is larger than the jump on the Full score (62.75 → 65.55), consistent with the training-split note above. The Full score is the one to compare. Not on JevBench or Artificial Analysis as of 17 Oct (we did not find a v1.2 row).

What is vendor-run

  • The card’s 68.09 / 67.75 and vision 77.43 / 72.53 are Perplexity’s runs. The board’s public figure for the 27B is 67.74.
  • 96 ms median for one request is Perplexity’s number on one GB200 with the fast path; the card gives no p95 for it. The “up to 5x faster” claim in the launch posts rests on the joint format, and we have not reproduced it.
  • ECE 0.014 is on a 1,497-question development set, with the temperature fitted in the joint format on separate data. On datasets held out from all training the card says ECE is about 0.05 and “probabilities remain somewhat over-confident”.

Fit / anti-fit

Fit: a hosted or self-hosted decision model with top-5 independent accuracy, images, and one cheap request for several facets of one decision. Flash fits high-volume routing and labelling where half the input price matters (Perplexity’s own advice is to keep Flash only if it passes your accuracy bar).

Anti-fit: ranking many independent candidates in one request, long inputs beyond ~32k tokens without your own validation, or a hard latency budget with no p95 published.

Limits

  • Needs the bundled code. Loading as a plain causal backbone gives different answers.
  • 32k training length; the card spot-checked up to ~243k tokens (about 95–107 GiB peak memory on a GB200).
  • Hosted limit: 10 requests per second, shared by both models; requests over the input limit are split into groups and the state is billed again for each group.
  • v1.1 clients: names pplx-decider-v1.1-27b and pplx-decider-v1-27b now resolve to v1.2.

Why it is in the catalog

Public Apache-2.0 weights, a documented typed API with prices, and an independent top-5 result on the official board. It supersedes the v1.1 page as the current Perplexity entry.

Score working (17 Oct 2026)

27B (headline).

  • Maturity = availability 8 × 0.30 + docs 8 × 0.25 + integrations 5 × 0.25 + support 3 × 0.20 = 2.40 + 2.00 + 1.25 + 0.60 = 6.25 → 6.3
  • Capability = types 10 × 0.30 + calibration 7 × 0.25 + latency 7 × 0.25 + accuracy 8 × 0.20 = 3.00 + 1.75 + 1.75 + 1.60 = 8.10 → 8.1
  • Adoption = engagement 30 × 0.35 + ecosystem 20 × 0.35 + production 2 × 0.30 = 10.50 + 7.00 + 0.60 = 18.10 → 18

Flash (4B), separate working.

  • Maturity 6.3: same subscores as the 27B (8 / 8 / 5 / 3).
  • Capability = 10 × 0.30 + 7 × 0.25 + 4 × 0.25 + 8 × 0.20 = 3.00 + 1.75 + 1.00 + 1.60 = 7.35 → 7.4. Latency 4: the card gives no Flash-specific number.
  • Adoption = 22 × 0.35 + 15 × 0.35 + 2 × 0.30 = 7.70 + 5.25 + 0.60 = 13.55 → 14.

Availability 8: Apache-2.0 weights plus a hosted API, but custom code is required (precedent: v1.1). Docs 8: card, API docs, cookbooks. Integrations 5: hosted API and community MLX / GGUF / NF4 conversions, no gateway listing found (precedent: v1.1). Support 3: no SLA, one rate tier.

Calibration 7: a published temperature and dev-set ECE, but ECE about 0.05 on held-out data (the card’s own words: still over-confident). Latency 7: median 96 ms on a GB200, vendor run, no p95. Accuracy 8: independent board result (Full 65.55, #4). The training-split caveat is a reason not to go to 9.

Adoption engagement 30 (v1.1 was 25): launch posts with 623 and 277 likes and 76k views on the main one, 19 + 16 HF likes in two days. Ecosystem 20 (unchanged): the Decisions API default, old names re-routed, MLX / GGUF / NF4 community builds. Production 2: no named users.