Daily News · 2 min read

Perplexity AI Updates: October 22, 2026

perplexity research distillation reinforcement-learning perplexity-computer local-ai apple-silicon enterprise sovereign-ai agents

1. Perplexity and NVIDIA Introduce JOLT, Training the Teacher and Student Together in Self-Distillation

Perplexity. Perplexity Research and NVIDIA published JOLT (Joint On-policy Learning and Teaching), an extension of on-policy self-distillation in which one model acts as both a teacher with privileged information (verified answers, passing solutions, verifier data) and a student without it. The student minimizes reverse KL to the teacher’s token scores, while the teacher is itself trained to give guidance that moves the student toward its own reward; a JOLT+ variant also adds the student’s outcome reward. Against GRPO, JOLT+ gained 4.2 points on MATH (Qwen3-0.6B), 3.0 on LiveCodeBench v6 (Qwen3-8B), 9.9 on AppWorld (Qwen3-4B) and 2.7 pass@1 on Terminal-Bench 2.1 (Qwen3-30B-A3B), and on MATH it reached GRPO-level accuracy with roughly 9.5 to 14.4x fewer completion tokens. The paper is on arXiv; code is promised “soon.” Source

2. Perplexity Computer Can Now Run Fully Local on Apple Silicon Macs

Perplexity. The Perplexity Mac app gained a fully local mode that keeps Computer’s planning, reasoning, execution and file processing on device, using Qwen 3.6 or PPLX Pro, a post-trained Qwen variant. Perplexity recommends 36 GB of memory or more; the mode works with local files and apps such as Notes, Messages, Mail and Numbers, runs code and tools in isolated sandboxes, and can optionally escalate to the cloud for browsing, connectors or 15+ frontier models. It is available to Pro, Max and Enterprise subscribers alongside the existing Hybrid Compute mode, and Perplexity pitches it as a way to control credit spend as well as keep private data local. Source

3. Valarian Partnership Puts Perplexity Computer on Air-Gapped Institutional Infrastructure

Perplexity. Perplexity partnered with UK sovereign infrastructure company Valarian, whose ACRA platform runs Perplexity Computer on servers or customer-controlled clouds that institutions operate, including air-gapped networks. Workloads are split into “enclaves” with infrastructure-enforced access to data, tools, users and credentials, audit records go to security teams, and institutions can run open-weight models such as Nemotron, GLM, Kimi, DeepSeek-R1 and Qwen. Deployments, aimed at government, finance, legal and healthcare, became available on October 20. Source