Release2026-10-12
Perplexity released pplx-decider-v1.1-27b. It lifts the causal mask and trains on more data. The Decision Index score rose from 56.4 to 61.56. It now outperforms Jev by over 3.5 points. The model uses a Qwen3.8-27B backbone. It requires noncausal attention mode for full performance. A separate readout head is used instead of a full vocabulary lm_head. Images must fit within an 8,192-token limit. Training capped images at 512×512 pixels. The model is not deployed by any inference provider yet.
What changed
pplx-decider-v1-27b used a causal mask and scored 56.4 on the Decision Index.
What it unlocks
Making decisions with higher accuracy using the same model backbone.
- 61.56 Decision Index score
- 56.4 to 61.56 Decision Index improvement
- 26B parameters
- 2,480 downloads last month
What you need to act on it
- Python 3.12+
- authenticated Hugging Face access
- CUDA GPU with ~49 GiB memory
- install pinned dependencies
Sources