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. Perplexity has released its solution for Mac this week. Hybrid compute A single can be split into two Perplexity Computer Task between frontier models on the cloud, and compact models on the Mac of the user. A privacy gate on the device decides what can cross the border. The classifier that powers the gate was also made available by Perplexity.
Is it deployable? Hybrid compute is available for Pro, Max and Enterprise subscriptions on Apple silicon Macs running macOS 15.x or higher with a minimum of 24GB unified memory. (32GB preferred). Installing the local model from within the Mac application is easy, as it requires no additional runtimes or API keys, neither Ollama nor separate runtimes.
How hybrid computing works
Design decision is made on the direction orchestration will take. Computer performs all tasks in the cloud. Frontier models are responsible for web searching, planning, and long horizon reasoning. Computer sends a task to a local Mac model when interacting with private or sensitive files. This is done without having to restart the job or lose context. The two halves are then combined into ONE result.
The inversion of the local compute Mode Perplexity was shipped with NVIDIA DGX Spark a few weeks earlier. It starts at the hardware of the user and can be escalated to cloud models if permission is granted. Same orchestrator, opposite default.
Computer is compatible with iPhone. This allows a task to be remotely triggered while the Mac on your desk executes sensitive steps. The Perplexity app positions an always-on Mac Mini as the dedicated node of local inference for this exact pattern.
Load-bearing components of the privacy gate include
An on-device classification software inspects the protected files before they are sent up to the cloud. After that, the gate can apply one of four options: Keep it local, Mask the sensitive sections, Refuse the Action, or Ask the User for Consent. The most sensitive data are credentials, government IDs and payment card numbers. The values that are masked out on their way to the cloud and replaced by stand-ins when they return, will be restored.
PII-Tracer This is a bidirectional 0.6B encoder, adapted from the Qwen3 Backbone. It replaces the casual mask with bidirectional attention that takes into account padding over a window of 4,096 tokens. One linear tagging device emits 37 labels – one label for outside the span plus BIOES labels for nine PIIs – and another head can predict whether a given conversation is sensitive. The training consisted of three epochs with approximately 714,000 samples. A constrained Viterbi encoder clarifies the sequence label at inference.
PII-TRACEThe benchmark contains 37,431 character level identifiers in 13 languages, 10 writing systems and 13,148 synthesized conversations. The central claim of the report is that it has found You can find out more about this by clicking here. In a lengthy conversation, PII is different from finding Each copy It is a fact.
PII Tracer has the highest character F1 (0.629) and is second in 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.
Interesting to note that single window recall falls from 0.975% for conversations less than 1000 characters down to 0.687 when the conversation is 10,000 or more characters. The fix for Perplexity is not retraining, but decoding: sliding windows with 50% overlap lift character recognition from 0.830 up to 0.965. Multi-mention detection goes from 0.794 up to 0.954.
One task can be split between cloud and Mac
function T(fn,ms){timers.push(setTimeout(fn,reduce?Math.min(ms,120):ms));}
Function ClearT(){timers.forEach(clearTimeout);timers=[];}
Resize the image using this function(){try{parent.postMessage({pplxHce:true,height:document.body.offsetHeight+40},’*’);}catch(e){}}
function chips(host,list,cls,step){
host.innerHTML=”;
list.forEach(function(t,i){
var c=document.createElement(‘span’);c.className=”chip “+cls;c.textContent=t;host.appendChild(c);
T(function(){c.classList.add(‘show’);resize();},step*(i+1));
});
}
Function paint(){
If s = S, then the value of var s will be:[sel];fnEl.textContent=s.fn;
txtEl.innerHTML=s.txt.replace(/
}
function reset(){
ClearT();busy=false;run.disabled=false;run.textContent=”Start task”;
cloud.innerHTML=”;mac.innerHTML=”;cnt.textContent=”idle”;
flow.classList.remove(‘go’);scan.classList.remove(‘go’);
ask.classList.remove(‘show’);led.classList.remove(‘show’);
Paint();resize();
}
Function Start(){
if(busy)return;clearT();busy=true;run.disabled=true;run.textContent=”Running…”;
cloud.innerHTML=”;mac.innerHTML=”;ask.classList.remove(‘show’);led.classList.remove(‘show’);paint();
S=S[sel];
flow.classList.add(‘go’);
chips(cloud,s.cloud,’cloudy’,450);
T(function(){chips(mac,[s.mac[0]],’local’,300);cnt.textContent=”scanning on device…”;scan.classList.add(‘go’);resize();},1500);
var marks=txtEl.querySelectorAll(‘mark’),n=0;
marks.forEach(function(m,i){T(function(){m.classList.add(‘hit’);n++;cnt.textContent=n+’ sensitive span’+(n>1?’s’:”)+’ found’;},2000+i*420);});
T(function(){
flow.classList.remove(‘go’);scan.classList.remove(‘go’);
askP.textContent=”Privacy Gate found “+s.fn+” in “+s.kinds+”. You can choose how to run this step.’;
ask.classList.add(‘show’);run.textContent=”Awaiting your call”;resize();
},2000+marks.length*420+500);
}
function decide(c){
ClearT();ask.classList.remove(‘show’);
S=S[sel];flow.classList.add(‘go’);
chips(mac,s.mac,’local’,350);
var v=root.querySelector(‘#hceVerdict’),o=root.querySelector(‘#hceOut’),
i=root.querySelector(‘#hceIn’),cr=root.querySelector(‘#hceCred’);
if(c===’local’){
v.textContent=”Step ran on your Mac. Cloud never saw the file.”;v.style.color=”#8FE3EC”;
o.className=”lv good”;o.textContent=”Anonymized research questions only”;
i.textContent=”Full file, every identifier, and the edited draft”;
cr.textContent=”Spent on cloud research steps only — local work uses none”;
T(function(){chips(cloud,[‘Answering anonymized questions’,’Merging both halves into one result’],’cloudy’,400);},600);
}else if(c===’mask’){
v.textContent=”Sent with stand-ins. Values restored on return.”;v.style.color=”var(–cyan)”;
o.className=”lv”;o.textContent=”File text with PERSON_1 / EMAIL_1 / ID_1 substituted”;
i.textContent=”The real values, and the mapping used to restore them”;
cr.textContent=”Spent on the cloud reasoning step”;
T(function(){chips(cloud,[‘Reasoning over masked text’,’Restoring values locally on return’],’cloudy’,400);},600);
}else{
v.textContent=”Uploaded as-is. The gate logged it and stepped aside.”;v.style.color=”var(–red)”;
o.className=”lv warn”;o.textContent=”The complete file, identifiers included”;
i.textContent=”Nothing withheld — an audit entry on Enterprise plans”;
cr.textContent=”Spent on the full cloud step”;
T(function(){chips(cloud,[‘Reasoning over the raw file’,’Identifiers now in cloud context’],’bad’,400);},600);
}
T(function(){
flow.classList.remove(‘go’);led.classList.add(‘show’);
busy=false;run.disabled=false;run.textContent=”Run again”;cnt.textContent=”decision applied”;resize();
},1700);
}
root.querySelector(‘#hcePills’).addEventListener(‘click’,function(e){
var b=e.target.closest(‘button[data-s]’);if(!b)return;
root.querySelectorAll(‘#hcePills .pill’).forEach(function(p){p.classList.remove(‘on’);});
b.classList.add(‘on’);sel=+b.dataset.s;reset();
});
ask.addEventListener(‘click’,function(e){var b=e.target.closest(‘button[data-c]’);if(b)decide(b.dataset.c);});
run.addEventListener(‘click’,start);
root.querySelector(‘#hceReset’).addEventListener(‘click’,reset);
Paint();resize();
window.addEventListener(‘load’,resize);
if(window.ResizeObserver){new ResizeObserver(resize).observe(document.body);}
})();
”
style=”width:100%;border:0;overflow:hidden;display:block;” height=”760″ scrolling=”no”
title=”Perplexity hybrid compute explainer” loading=”lazy”>
Models and controls are available
Perplexity lists three models as local at launch, including Gemma4 E4B (also known as Gemma 4), Qwen3.635B-A3B (also known by the model name Perplexity), and one that has been post-trained to work with Computer. Product page setup flows point to the one-click installation of Qwen 3.827B by PPLX; Perplexity’s Hugging Face org Matching carries pplx-computer-qwen-3-8-27b Buildings along the way pplx-pii-masking-vllmThe 0.6B token classification model is 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. It is this piece which makes the product usable by legal, health care, and finance teams, rather than being just an interesting feature.
What you need to know
- Hybrid computing starts tasks on the cloud, and then transfers sensitive data to the local Mac model in mid-task.
- A PII classification gate on the device gates boundaries: local only, mask, refuse or ask.
- In terms of character F1 (0.629) and finding all recurring references (79.4%), PIITracer (0.6B), the leading detector among 12 other detectors, is PII-Tracer.
- The long-context retrieval drops below 0.965 after 10K characters.
- For Pro, Max and Enterprise, the system must run on Apple hardware, macOS 15+ with a minimum of 24GB unified RAM.
Click here to find out more Perplexity announcement. PII-TRACE research postThe. model on Hugging Face. Also, feel free to follow us on Twitter Join our Facebook group! 150k+ML SubReddit Subscribe now our Newsletter. Wait! Are you using Telegram? now you can join us on telegram as well.
Want to promote your GitHub repo, Hugging Face page, Product release or Webinar?? Connect with us
Asif Razzaq serves as the CEO at Marktechpost Media Inc. As an entrepreneur, Asif has a passion for harnessing Artificial Intelligence to benefit society. Marktechpost is his latest venture, a media platform that focuses on Artificial Intelligence. It is known for providing in-depth news coverage about machine learning, deep learning, and other topics. The content is technically accurate and easy to understand by an audience of all backgrounds. This platform has over 2,000,000 monthly views which shows its popularity.

