pplx-decider-v1.1-27b
Perplexity's Apache-2.0 multimodal decision model on Qwen3.8-27B, reporting 61.56 on the Decision Index versus 57.9 for Jev.
Overview
pplx-decider-v1.1-27b is Perplexity's update to pplx-decider-v1-27b, a multimodal decision model fine-tuned from Qwen3.8-27B. The model card reports the Decision Index rising from 56.4 to 61.56, more than 3.5 points above Jev's 57.9, attributed mainly to lifting the causal mask on full-attention layers and training on more data including tasksource. By category it leads on language, retrieval and arts, trails Jev slightly on knowledge, and is near v1 on tools.
The checkpoint uses a native decision layout: a Qwen backbone plus a separate BF16 readout head over 255 decision slots rather than a full-vocabulary LM head. The included inference code preserves the noncausal attention mode and applies the saved calibration temperature; serving stacks need a special export. It needs a CUDA GPU with room for about 49 GiB of weights. Per the list summary it is Apache-2.0 and also served through the Perplexity Decisions API and OpenRouter.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which of up to 255 candidates fits the state? | — |
| Yes / no | Is this statement true of the state?(inferred) | yes · no |
How it works
- Download the checkpoint and install pinned requirements.
- Load with the included DecisionModel.
- Call predict, which applies calibration temperature.
- Or use the Perplexity Decisions API or OpenRouter.
Strengths
- Reported to exceed Jev on the Decision Index.
- Open Apache-2.0 weights.
- Hosted options available.
Caveats
- Needs about 49 GiB of GPU memory for weights.
- Requires the included inference code to reproduce results.
Getting started
Use the Perplexity Decisions API or OpenRouter, or download from Hugging Face and use the included DecisionModel.predict on a large CUDA GPU.At a glance
- Type
- Service
- Status
- production
- Language
- Python
- Runtime
- PyTorch on GPU, hosted API
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Perplexity's Apache-2.0 multimodal decision model fine-tuned from Qwen3.8-27B; v1.1 lifts the causal mask and trains on more data, and its card reports 61.56 on the Jev Decision Index against 57.9 for Jev. Also served through the [Perplexity Decisions API](https://docs.perplexity.ai/docs/decisions/quickstart) and [OpenRouter](https://openrouter.ai/perplexity/pplx-decider-v1.1-27b).
AnotiaWang/awesome-decision-models
Frequently asked questions
How does it compare with Jev?
The card reports 61.56 vs 57.9 on the Decision Index overall.
What changed from v1?
Lifting the causal mask and training on more data.
Can I use it without a GPU?
Via the Perplexity Decisions API or OpenRouter.
Related projects
Small open vision-language decision models (2B, 4B, 9B) that answer typed questions about photos and records with a trained can't-tell.
ModelbetaPythonHead-to-head eval page where TrustedRouter's gev-1.0 outperforms Jev on the same JevBench questions and also accepts images.
Articlediscussion- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython Free hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaComfyUI nodes for Jev-style noul, choice and score decisions over images and text using a local imajev adapter on Qwen3.5.
ProjectbetaPython33-chapter Chinese tutorial on agentic RL and GRPO, including a chapter on when Jev-style external decision models beat training with RL.
ProjectresearchPython