Agentic assistants have a structural problem: the context that makes them useful — deal documents, privileged files, client records — is exactly the context users cannot send to a cloud endpoint. This week, Perplexity shipped its answer for Mac. Hybrid compute splits a single Perplexity Computer task between frontier models in the cloud and a compact model on the user’s Mac, with an on-device privacy gate deciding what may cross the boundary. Perplexity also open-sourced the classifier behind that gate.
Is it deployable? Yes, hybrid compute is live for Pro, Max, and Enterprise subscribers on any Apple silicon Mac running macOS 15 or later with at least 24GB of unified memory (32GB recommended). The local model installs in one click from the Mac app, with no Ollama, no separate runtime, and no API key, and local work consumes no cloud credits.
What hybrid compute actually does
The direction of orchestration is the design decision. Computer starts every task in the cloud, where frontier models handle web search, planning, and long-horizon reasoning. When a step touches private files or sensitive data, Computer hands that step down to the local model on the Mac without restarting the task or losing context then merges both halves into one result.
This inverts the local compute mode Perplexity shipped on NVIDIA DGX Spark a week earlier, which starts on the user’s hardware and escalates up to cloud models with permission. Same orchestrator, opposite default.
Because Computer works with iPhone, a task can be triggered remotely while sensitive steps execute on the Mac at the desk. Perplexity positions an always-on Mac mini as a dedicated local inference node for exactly this pattern.
The privacy gate is the load-bearing component
Before anything from a protected file reaches the cloud, an on-device classifier inspects it and the gate applies one of four outcomes: keep it local, mask the sensitive spans, refuse the action, or ask the user for consent. Credentials, payment card numbers, and government IDs get the strictest handling. Masked values are swapped for stand-ins on the way out and restored when the cloud answer returns.
PII-Tracer is a 0.6B bidirectional encoder adapted from a Qwen3 backbone, replacing the causal mask with padding-aware bidirectional attention over a 4,096-token window. A linear tagging head emits 37 labels, one outside-span label plus BIOES position labels for each of nine PII types, and an auxiliary head predicts whether a conversation contains sensitive material. Training ran three epochs on roughly 714,000 samples; a constrained Viterbi decoder resolves the label sequence at inference.
PII-TRACE, the accompanying benchmark, contains 13,148 synthetic conversations across 13 languages and 10 writing systems, with 37,431 character-level identifier mentions. Its central claim is that finding most PII in a long conversation is not the same as finding every copy of it.
On results: across 12 detectors, PII-Tracer records the highest character F1 (0.629) and the second-best span-overlap and span-containment F1, behind GPT-5.6-sol. On consistency it leads by a wide margin — every mention found for 79.4% of recurring identifiers and 77.6% of cross-turn identifiers, versus 57.0% and 55.1% for GPT-5.6-sol. In the hardest bucket (6–10 mentions) it scores 0.691 against 0.464 for GPT-5.6-sol, 0.073 for GLiNER2-PII, and 0.045 for Claude Opus 4.8.
It is very interesting to know that single-window recall drops from 0.975 on conversations under 1,000 characters to 0.687 at 10,000 characters or more. Perplexity’s fix is decoding, not retraining: 50%-overlap sliding windows lift overall character recall from 0.830 to 0.965 and multi-mention consistent detection from 0.794 to 0.954 on the same checkpoint.
Explainer: how one task splits across cloud and Mac
Models, controls, and availability
Perplexity’s announcement lists three local models at launch: Gemma 4 E4B, Qwen3.6 35B-A3B, and a Perplexity model post-trained for Computer. The product page’s setup flow points to a one-click download of PPLX Qwen 3.8 27B; Perplexity’s Hugging Face org carries matching pplx-computer-qwen-3-8-27b builds alongside pplx-pii-masking-vllm, the 0.6B token-classification model behind the gate.
For Enterprise, admins can set org-wide rules for what must stay on device, what may be masked, and what requires explicit approval — plus audit logs for when information leaves a machine. That is the piece that makes this usable for legal, healthcare, and financial teams rather than just interesting.
Key Takeaways
- Hybrid compute starts tasks in the cloud and hands sensitive steps down to a local model on the Mac, mid-task.
- An on-device PII classifier gates the boundary: keep local, mask, refuse, or ask.
- PII-Tracer (0.6B) leads 12 detectors on character F1 (0.629) and on finding every recurring mention (79.4%).
- Long-context recall drops to 0.687 past 10K characters; sliding-window decoding recovers it to 0.965.
- Runs on Apple silicon, macOS 15+, 24GB unified memory minimum, for Pro, Max, and Enterprise.
Check out the Perplexity announcement, the PII-TRACE research post, and the model on Hugging Face. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
原始发布方：MarkTechPost（RSS）
原文时间：2026-09-02 13:12:00 +08:00
