Perplexity Decider v1.1 License: Apache-2.0 Weights, Not Private

What happened

When Perplexity shipped the second version of its decision model, pplx-decider-v1.1-27b, one early account raised a concern. According to that report, the v1.1 repository appeared as "private" and showed no license line, even though the launch announcement called the model open 5. For a company that had built its pitch on open weights, that was a fair question to ask.

The available evidence now points one way. Two independent community conversions of v1.1 published on Hugging Face both state that the original model is licensed under Apache-2.0. Both say they carry over the upstream LICENSE and NOTICE files. The mlx-community 4-bit build for Apple Silicon says it uses "the same license as the original model" and credits Perplexity fully 2. A separate bitsandbytes NF4 quantization by lmcoleman says the base model's Apache 2.0 license applies to the derivative. It also adds an addendum to NOTICE listing its changes and names the exact upstream revision it was quantized from 3.

These are third-party statements, not a fresh declaration from Perplexity. Still, they are specific and consistent. Two separate repos would be unlikely to cite and reproduce license files from a gated upstream that had none. Our reading is that the "private" observation was most likely a snapshot taken during the release window, not evidence that the weights are proprietary.

The model and its fast release cycle

The licensing question sits inside a very fast release schedule. On October 1, 2026, Perplexity opened a Decisions API backed by pplx-decider-v1-27b. That model is an open-weight fine-tune of Qwen3.8-27B released on Hugging Face under Apache 2.0, with a 262,144-token context window 4. The API returns answers in three typed formats: yes/no, multiple choice, and ordered score. It launched at $0.04 per million input tokens, with output free. Perplexity's documentation shows responses arriving in under two seconds for short prompts and taking up to 23 seconds near the input ceiling 4.

Five days later, on October 6, v1.1 arrived. It cut the API price in half to $0.02 per million input tokens 5.

It helps to understand what these models are. The mlx-community card states plainly that this is a classifier, not a chat model 2. You supply a state and a question. The state can be text, JSON, chat messages, or up to eight images. A single forward pass returns calibrated probabilities over the answer options, with no generation step 2. That design helps explain the free-output pricing: there is very little output to bill for.

Benchmarks: vendor numbers versus third-party leaderboards

The two releases came with different kinds of evidence. For v1, Perplexity relied on its own panel of eleven benchmarks and 7,210 samples. On that panel the model scored 85.71% overall against 84.51% for TypeSafe's Jev. The widest gap was on RAGTruth, at 88.80% versus 77.27% 4.

For v1.1, the headline figure came from outside the company. The model topped the newly launched Hugging Face Decision Index 0.3 with a score of 61.56, compared with Jev's 57.9 5. CEO Aravind Srinivas also claimed on October 9 that the model leads DecisionBench at the lowest cost in its category 5.

The Decision Index figure has already reached ranking roundups. One October list places pplx-decider v1.1 second overall and calls it the strongest open option, noting it is multimodal and available as both open weights and an API 1. The same list ranks OpenAI's hosted Decisions API, running gpt-6-luna and still in public beta, third. It prices that service at $0.10 per million input tokens and recommends it for single-vendor and compliance-focused shops 1.

One caveat applies to anyone using the quantized builds. The NF4 maintainer notes that the Decision Index has not been run on the compressed weights, and that the derivative inherits any biases in the base model 3. Leaderboard scores for the full-precision model do not automatically carry over to 4-bit versions.

Why it matters

The license question matters because Apache-2.0 is central to the model's competitive position. Against a hosted beta that costs five times as much per input token 1, Perplexity's advantage depends on teams being able to self-host, quantize, and redistribute the weights. The community ports are already doing this on Apple Silicon and AMD hardware 23. If v1.1 had quietly become private, that advantage would have disappeared.

The episode also says something about this market. Decision models are moving on weekly cycles. Credibility is shifting away from vendor self-benchmarks and toward price and third-party leaderboards 5. In that environment, license metadata needs the same scrutiny as benchmark tables.

The flag about the private repo was worth raising. The derivative chain now suggests it was a false alarm. Even so, the most reliable confirmation would be a license line on Perplexity's own model card.