AutoJev-27B

Last updated:

Open—$0.02/M input

Quadrant scores

See the full quadrant

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

  • Maturity6.6/10

    Apache-2.0 weights and MIT code with a /v1/systemone server, and since 01/10 the same weights as a hosted, priced Perplexity API ($0.02/M input since 6 Oct) with quickstart, OpenAPI reference, rate limits and a cookbook. The hosted endpoint is a preview (gateway-preview v0.1.0) on its own path, with no SLA.

  • Capability7.6/10

    Official JevBench (17 Oct): Intelligence 55.6, calibration 87.6, p50 0.097 s. Noul, Choice and Score in one pass. Temperature scaling is documented. No latency figures published.

  • Adoption24/100

    Perplexity's launch post reached ~540 likes and ~116k views within hours, and Perplexity now serves the model. Before that: ~47 HF likes, ~1k downloads, 110 GitHub stars and a community GGUF. No named production users yet.

Vendor claims

News covering this model

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 Perplexity’s Decisions API serves these exact weights. perplexity-ai/pplx-decider-v1-27b (Hugging Face, 01/10 18:24 BRT, Apache-2.0) has byte-identical files to denis-pplx/autojev-27b: the same SHA-256 for all 11 safetensors shards, readout.safetensors, config.json and decision_config.json. Its NOTICE calls it the “AutoJev 27B curated SFT checkpoint” and it bundles the autojev source. So the hosted pplx-decider-v1-27b is this catalog entry, not a new model. Perplexity has not said so in its launch post; the hash comparison is ours (01/10).

AutoJev-27B is an open decision model published by the Hugging Face / GitHub user denis-pplx on 19 September 2026. It is a full-weight fine-tune of Qwen/Qwen3.8-27B that returns probabilities over the options you supply, in one forward pass per question. The repo ships a server that speaks POST /v1/systemone with choice, noul and score, a browser playground and optional base64 images.

The author describes it as an “independent implementation inspired by Jev”. It is not a TypeSafe product and is not trained on Jev outputs as far as the card says. It is also unrelated to AutoTrust’s JEV models (now cataloged as GEV-26B-Decide), which use a similar-sounding name.

v1.1 is a different model (6 Oct 2026). Perplexity’s pplx-decider-v1.1-27b is a new training run on the same Qwen3.8-27B backbone (noncausal attention, more data) and is now the API’s default model, with its own page. This page stays on v1 = AutoJev-27B, which the API still serves as pplx-decider-v1-27b, now at $0.02/M input.

Perplexity Decisions API (1 October 2026)

@perplexitydevs launched a Decisions API on 1 October 2026 (15:24 BRT; ~540 likes and ~116k views by the evening).

Attribute Value
Endpoint POST https://api.perplexity.ai/v1/decisions, model pplx-decider-v1-27b
Body System One shape (state, questions) with noul, choice, score and images, on its own path (not /v1/systemone)
Price $0.02 / M input tokens since 6 Oct 2026 (launched at $0.04); output tokens free; no per-request fee
Limits 10 requests/s per organization; ~250k-token context per the launch post
Timing (Perplexity’s tests, 30/09) a few hundred input tokens under 2 s; ~90k tokens 5 s; ~190k tokens 14 s
Status OpenAPI labelled “gateway-preview” v0.1.0
Weights perplexity-ai/pplx-decider-v1-27b, identical to AutoJev-27B

Perplexity’s chart (UnverifiedClaim): 85.71% overall vs Jev 84.51% on an 11-benchmark panel, 7,210 rows run through its own API; Jev is ahead on WinoGrande, BBH, JevBench hard, TruthfulQA, Belebele and JudgeBench. The cookbook shows ticket triage.

The company field now reads “denis-pplx · hosted by Perplexity”. Perplexity’s post calls it “our multimodal decision model” without mentioning AutoJev; the author’s handle suggests he works at Perplexity, but neither side states it.

Specs

Attribute Value
Author denis-pplx (HF + GitHub)
Release 19 Sep 2026 (HF repo 16:54 BRT; GitHub repo created the same day)
Base model Qwen/Qwen3.8-27B
Parameters 26.1B (BF16, ~49 GiB of weights)
Training Full-weight SFT on one H200, 73,000 unique examples, 286 updates; checkpoint 200 released; scalar temperature fitted separately
Training data Generated and curated by the author’s agent pipeline; the exact corpus is not bundled
API POST /v1/systemone (choice, noul, score, optional images); FastAPI docs at /docs
License Weights Apache-2.0, code MIT
Latency No figures published (0–0ms in the frontmatter means “no data”)

Decision Index 0.2.1 (official Space)

Measured by the official Decision Index Space (data snapshot 28 Sep 2026, 71 systems):

System Index (balanced skill) Rank
Jev 1.13 57.91 #1
Surogate Rune 26B-A4B v3 57.44 #2
AutoJev-27B 56.40 #4
simple-jev (Qwen3.8-27B, runtime) 55.74 #5

The index also measures calibration: for AutoJev-27B, accuracy 0.730, mean confidence 0.738, ECE 0.018 and Brier 0.361. That is tighter than the author’s own ECE figure.

Note the gap to simple-jev on the same Qwen3.8-27B base: 0.66 points. Most of AutoJev’s accuracy comes from the base model; the fine-tune adds a trained decision readout and better calibration.

Author’s own results (UnverifiedClaim)

Model Overall accuracy ECE Brier
Qwen3.8-27B 69.83% 0.0648 0.408
AutoJev-27B 84.60% 0.0428 0.220
Jev 82.79% 0.0527 0.254

These come from the author’s curated held-out set (holdout-v3), which is not public. Beating Jev on your own test set is a different claim from matching Jev across public suites; on the official index AutoJev sits 1.5 points below Jev.

Community GGUF

thomasgauthier/autojev-27b-GGUF ships a Q4_K_M quant and a vision projector (~3k downloads). It does not run on mainline llama.cpp or LM Studio: you need the jev.cpp fork, which adds the AutoJev classifier head and a --system-one server flag. Another copy, arafathusayn/autojev-27b, is a plain re-upload.

Fit / anti-fit

Fit when you need: a local, TypeSafe-compatible endpoint with good measured calibration, agent gating and routing decisions, or a base to study (training code is public).

Anti-fit when you need: a small GPU (about 49 GiB of BF16 weights plus overhead), published latency, natural-image accuracy (the author says image support is not an accuracy claim), or the training corpus.

Limits

  • The training corpus is not published; the headline accuracy is on a private held-out set.
  • The card still says “while the weights are private, authenticate”; the repo was public and ungated when we checked on 30 Sep 2026.
  • “Built by autonomous agents” describes the author’s process; it does not change how the model is evaluated here.
  • No p50/p99 latency numbers; Perplexity gives only coarse timings by input size.

Why it is in the catalog

The weights were fully fine-tuned for typed decisions and ship with an open System One-compatible server that returns probabilities. That passes the catalog criteria. The official index places it among the top open entries.

JevBench v1.6.1 (Benchmark Heaven, checked 7 Oct 2026)

On the v1.6.1 open-weights board (added 7 Oct; not comparable with v1.5.7), AutoJev-27B is #46 at 18.5. Its capability is 71.6 (Intelligence 55.6, Calibration 87.6, raw p50 0.34 s), but self-hosting costs $0.23 per 1,000 decisions (7.0× Jev), so it sits outside the Jev-class caps. Decision Index 0.3 dropped pplx-decider v1 (these weights) as superseded by v1.1. See news. No catalog score change.

JevBench v1.5.7 (earlier edition, checked 5 Oct 2026; not comparable with v1.6)

AutoJev-27B is #55 at official 19.5 on v1.5.7. See news. No catalog score change.

Board audit (17 Oct 2026)

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) AutoJev-27B Capability 71.6; Intelligence 55.6; calibration 87.6; p50 0.097 s; $0.23/1k
ImageJevBench v0.3.0 AutoJev-27B Capability 62.1; Intelligence 37.6; calibration 86.7; p50 0.48 s; $0.054/1k
ImageJevBench v0.1.5 archive (frozen) AutoJev-27B Capability 75.9; Intelligence 69.3; calibration 82.4; p50 0.43 s; $0.049/1k

Score update (audit, 17 Oct 2026). Capability 7.2 → 7.6 = types × 0.30 + calibration × 0.25 + latency × 0.25 + accuracy × 0.20 = 8 × 0.30 + 8 × 0.25 + 7 × 0.25 + 7 × 0.20 = 7.55 → 7.6. Types, calibration and latency come from the earlier working (8 / 8 / 7); only accuracy is re-derived from the board (7; was 5): JevBench Intelligence 55.6 ÷ 63.6 × 8 = 7.0, half-up. Maturity and adoption unchanged.

Hugging Face on 17 Oct (denis-pplx/autojev-27b): 3,056 downloads, 52 likes, licence tag apache-2.0, last modified 2026-09-22.