Perplexity Decisions API: 27B Model's Benchmark Edge Is Narrow

What Perplexity shipped

On October 1, 2026, Perplexity launched a Decisions API and published the weights of the model behind it, pplx-decider-v1-27b. 15 The model is a fine-tune of Qwen3.8-27B, released under the Apache 2.0 license, with a 262K-token context window. 12 The hosted API costs $0.04 per million input tokens, and output is free. 1

The main idea is simple. The model does not generate text. A developer sends in content along with a typed schema, such as a yes/no question, a multiple-choice selection or a score. The model returns a calibrated probability for each allowed option. 1 Perplexity's pitch is that this removes the brittle parts of LLM-based classification: no parsing, no regex, and no repairing malformed JSON. 1 Several outlets describe it as multimodal, meaning it can weigh images as well as text when making a call. 45

The Hugging Face artifacts show this design clearly. A community 4-bit EXL3 quantization notes that the model skips the usual language-model head for its final answer. Instead, it produces a hidden state and passes it through a separate decision head, stored as a readout.safetensors file. Running it therefore requires a custom script rather than standard generation code. 3 The repository also lists the architecture as Qwen3_5Model. That label sits a little oddly next to the "Qwen3.8" base cited elsewhere, a small discrepancy worth watching as documentation matures. 13

The benchmark claim, read closely

Perplexity's headline number is 85.71% across an 11-benchmark panel. That compares with 84.51% for Jev, TypeSafe's decision model, and 74.76% for the untuned Qwen base. 12 The gain over the base model is substantial and suggests the fine-tune does real work.

The lead over Jev is a different matter. It amounts to about 1.2 points on an aggregate, and as one observer pointed out, Jev still wins six of the eleven individual benchmarks. 2 Perplexity therefore takes the overall average while losing the majority of head-to-head tasks. That pattern usually means a few large margins are lifting the mean. One analysis calls the result "a strong same-tier rival, not a coronation" and notes that no independent rerun exists yet. 1 Every number cited so far comes from the vendor's own panel. 12

Coverage of the release has been uneven. Some outlets framed it around the API and the Jev comparison. 15 One early writeup described a quiet weights drop with little accompanying documentation or benchmark data, positioning the model mainly as an internal routing component made public. 4 One account also notes the weights were already sitting on Hugging Face before the announcement landed. 2 Taken together, this suggests the benchmark story was assembled around a model that had first surfaced with little context.

A crowded launch day

Timing makes the comparison harder to judge. Perplexity's release came hours after Cloudflare announced Clef. 5 By one count, three vendors launched decision models on the same day. 1 That points to a new product category forming quickly: models that sit upstream in a pipeline and decide where a request should go, rather than answering it. 14

Routing and classification are unglamorous jobs that most labs keep in-house. 4 Releasing a production-grade router under a permissive license gives developers something to study, adapt and self-host. 4

Practicality

On the deployment side, the early signals are encouraging. The 4-bit EXL3 conversion loaded in about 3 seconds on an RTX 4090 with a warm file cache. It used roughly 12.6 GiB of resident memory and completed a first 120-token inference in about 1.5 seconds, with later requests running faster. 3 These figures come from a single community conversion, not an official benchmark. Still, they suggest a 27B decider can run on one consumer GPU, which matters for teams that don't want routing traffic leaving their own infrastructure.

The reading

The substance here is the format, not the leaderboard. Returning calibrated probabilities over a fixed schema is a cleaner contract for routing and moderation than coaxing structured text out of a chat model. An open-weight, Apache-licensed implementation that runs on one GPU is useful regardless of where it ranks.

The "edges Jev" framing deserves skepticism. A roughly one-point aggregate lead on a vendor-chosen panel, with the rival winning most individual tests, is effectively parity. Teams choosing a decider should run both models against their own label sets and check calibration on their own data rather than rely on the 85.71% figure. Until independent evaluations appear, the fair verdict is that Perplexity has entered a new category credibly, not that it has won it.