pplx-decider v1.1 27B
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity6.3/10
Apache-2.0 weights (Qwen3.8-27B + 255-row readout) plus a priced hosted API at $0.02/M input with quickstart, API reference, cookbook and 10 req/s per org. Needs the bundled code (noncausal attention), not a stock Transformers class. No SLA.
- Capability8.1/10
Official JevBench (17 Oct): Intelligence 75.7, calibration 89.9, p50 0.122 s. Noul / Choice / Score, images.
- Adoption16/100
Perplexity serves it as the default Decisions API model and the CEO amplified the launch. On Hugging Face it had ~170 downloads and 34 likes two days in. No named production users yet.
Vendor claims
- Decision Index score 61.56 vs Jev 57.9 and v1 56.4 (Perplexity's own run of Decision Index 0.2.1); ahead of Jev on Language, Retrieval, Tools and Arts, behind on Knowledge (48.18 vs 51.4)[Vendor claim — not independently verified]
- 'It scores the highest on the new Hugging Face Decision Index 0.3 benchmark' (true on the official board, see page)[Vendor claim — not independently verified]
- Most of the gain comes from lifting the causal mask and training on more data, including tasksource[Vendor claim — not independently verified]
publicado (OK do Fabricio 17/10 BRT). Board audit of 17 Oct 2026: see the “Board audit” section below; where it conflicts with older board statements on this page, it wins.
Read this first
- New weights, not a re-upload. v1 (
pplx-decider-v1-27b) was byte-identical to AutoJev-27B. v1.1 is a new training run on the same Qwen3.8-27B backbone, so it gets its own page. AutoJev keeps v1.- Two different index numbers. The card’s 61.56 is Perplexity’s own run of Decision Index 0.2.1. The official Decision Index 0.3 board (6 Oct 2026) scores it 62.75, rank 1. Do not mix them.
- You must use the bundled code. Full-attention layers run noncausal. Default causal inference “does not reproduce this checkpoint’s evaluated setup”, and there is no full-vocabulary
lm_head. The card still calls the repo “private”, but it is public.- Superseded on 15–16 Oct 2026. Perplexity released pplx-decider v1.2 and Flash v1.2 (joint requests, Full 65.55 on the official board, #4). The board no longer lists v1.1, and the hosted API routes the old
pplx-decider-v1.1-27bname to v1.2. This page keeps the v1.1 facts as written on 14 Oct; the “#1” below is from the 6 Oct board.
What it is
An open decision checkpoint from Perplexity. A Qwen3_5Model backbone (Qwen3.8-27B, 26.1B parameters in safetensors) plus readout.safetensors, a BF16 [255, 5120] decision head. The bundled DecisionModel.predict applies a saved calibration temperature. It is also the default model behind Perplexity’s hosted Decisions API.
Specs
| Attribute | Value |
|---|---|
| Weights | perplexity-ai/pplx-decider-v1.1-27b, Apache-2.0 |
| Published | Hugging Face ; announced 6 Oct 2026, 18:11 BRT |
| Base | Qwen/Qwen3.8-27B, fine-tuned; noncausal full-attention layers |
| Head | 255 decision slots, BF16 |
| Memory | About 49 GiB of weights plus working memory (CUDA) |
| Hosted API | POST https://api.perplexity.ai/v1/decisions, model pplx-decider-v1.1-27b; 1–128 questions; input under 262,144 tokens; 10 req/s per org |
| Price | $0.02 per 1M input tokens, output free (half of v1’s launch price; v1 stays selectable) |
| Question types | noul, choice, score; text, JSON or images as state |
Independent results (verified 11 Oct 2026)
| Board | Result |
|---|---|
| Decision Index 0.3 (official, 6 Oct 2026) | 62.75, rank 1 (not tied; tie band 0.9). Public 62.25 · same skills 61.10 · new domains 55.57. Jev 60.11 (rank 2 tie group). v1 left the board as superseded |
| JevBench v1.6.1 open-weights board (added 7 Oct, v1.7.11) | Composite 20.6 (#40). Intelligence 75.7 (highest open-weights row), Calibration 89.9, Speed 86.8, Cost 29.9. Capability 82.8 but outside Jev-class: $0.22 per 1,000 decisions = 6.73× Jev self-hosted. Raw p50 0.39 s |
| EdgeEvals System One preview (11 Oct 2026) | 643/669 credited; rank 1 tie in the current 35-row board, one severe error, 362 ms median. API cost: $0.02543/1,000 decisions, about $0.017 for 669 |
The low JevBench composite is a cost effect: the board prices self-hosting a 27B model on rented GPUs. It is not an accuracy result. The hosted price ($0.02/M) is a vendor decision and is not what the board measures.
EdgeEvals’ preview cases are authored and have not been reviewed by clinicians. This is a benchmark result, not clinical validation. The board gives pplx-decider v1.1 and GPT-6 Luna rank 1 even though their credited counts differ (643 vs 637); Mercury Decide has 630 and Jev 1.13 has 628, both at rank 2. Pplx-decider and Jev each have one severe error; GPT-6 Luna has none and Mercury has two. The current JSON has 35 rows, while the 10 Oct Microsoft snapshot recorded 38. The roster change is unexplained, so those snapshots should not be read as a simple eight-model net increase. This result does not change the catalog score.
Fit / anti-fit
Fit when you need: the strongest open decision checkpoint on the official index, a cheap hosted fallback with the same weights, images, and up to 128 questions per call.
Anti-fit when you need: a stock Transformers or vLLM class, a small GPU, the TypeSafe /v1/systemone path (Perplexity uses /v1/decisions with a System One-shaped body), or an SLA.
Why it is in the catalog
New decision-trained weights with an open license, a documented readout and calibration, a priced API, and the top official Decision Index 0.3 score.
Score working (7 Oct 2026)
- 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 25 × 0.35 + ecosystem 20 × 0.35 + production 2 × 0.30 = 8.75 + 7.00 + 0.60 = 16.35 → 16
Latency stays at 7: the board’s 0.39 s is self-hosted on a large GPU, and Perplexity publishes no hosted latency for v1.1.
Artificial Analysis Decision Index v1.0 (13 Oct 2026)
Per the AA launch article, Decider v1.1 27B is the cheapest at $17 per 1M tasks and on the cost Pareto frontier; WildGuardTest 94 (level with Jev 1.13). The article text doesn’t state its index score, so we don’t quote one. Scores unchanged. AA is a separate, independent board (four evaluations: AA-Omniscience-Decision, HLE-MultiChoice, When2Call, WildGuardTest). Its scores and ranks are not comparable with Decision Index 0.3 or JevBench. AA finds every decision model overconfident overall, mainly on the knowledge and reasoning tasks; most are fairly well calibrated on When2Call and WildGuardTest.
Board audit (17 Oct 2026)
publicado (OK do Fabricio 17/10 BRT). Checked against the official boards on 17 Oct 2026: Decision Index 0.3.2 (data of 16 Oct, 19:00 UTC), JevBench and ImageJevBench live. The boards are separate sources and are never averaged. Older board statements and rankings further down this page are superseded where they conflict with this section.
| Board | Row | Result |
|---|---|---|
| JevBench open weights (live 17 Oct) | Perplexity Decider v1.1 27B | Capability 82.8; Intelligence 75.7; calibration 89.9; p50 0.122 s; $0.22/1k |
Board-status statements on this page to recheck against the table (may be superseded):
- “Two different index numbers. The card’s 61.56 is Perplexity’s own run of Decision Index 0.2.1.”
Score update (audit, 17 Oct 2026). Capability 8.1 → 8.1 (unchanged) = types × 0.30 + calibration × 0.25 + latency × 0.25 + accuracy × 0.20 = 10 × 0.30 + 7 × 0.25 + 7 × 0.25 + 8 × 0.20 = 8.10 → 8.1. Types, calibration and latency come from the earlier working (10 / 7 / 7); only accuracy is re-derived from the board (8; was 8): JevBench Intelligence 75.7 ÷ 63.6 × 8 = 9.5, half-up, capped at 8. Maturity and adoption unchanged.
Hugging Face on 17 Oct (perplexity-ai/pplx-decider-v1.1-27b): 4,095 downloads, 155 likes, licence tag apache-2.0, last modified 2026-10-12.