Perplexity ships pplx-decider v1.1: bidirectional attention adds +2.8 DI, new sizes 27B/4B
antoine_chaffin · x · 2026-10-17
Perplexity's team announced the next iteration of its decider models in two sizes, 27B and 4B, plus a new attention schema for better batching.
v1.1's gains over v1 came from two changes: better training data (+2.4 DI) and switching from causal to fully bidirectional attention (+2.8 DI). But bidirectional attention creates a serving problem: the state attends to the question, so the state cache can't be reused and every new question re-encodes everything — cost scales as N × (state + question) instead of state + N × question. The rest of the thread tackles this caching challenge.
More from Models
- Perplexity Launches Decisions API: Decision Models at $0.01 per Million Input Tokens — AravSrinivas · 2026-10-17
- Arena weekly: Gemini 4 Argon hits #5 in Agent Arena, Meta Muse Wonder debuts #2 — arena · 2026-10-17
- Tesla's FSD is RL-trained to be polite: yields to turns, stops for pedestrians, avoids animals — ChrisGPT · 2026-10-17
- DeepSeek launches standalone Search & Fetch APIs at ~$0.70 per 1,000 successful requests, failures free — teortaxesTex · 2026-10-17
- Anti-hype distributed systems engineer: Claude now writes better code than me — IgorCarron · 2026-10-17
- Dev tries the new Jev model for on-device video detection and scene classification on iOS — amos_gyamfi · 2026-10-17