Perplexity's local agent runs on Alibaba's Qwen, not Claude or GPT
Two stories this morning about where the work actually happens — and who owns the model underneath it.
Perplexity's on-device agent, Portable Computer, runs on Alibaba's open-weight Qwen3.8-27B rather than Anthropic's Claude or OpenAI's GPT. The company built its own optimization layer — a post-trained "PPLX 27B" variant of Qwen, which it has published on Hugging Face — to handle local file processing, data analysis and coding on Nvidia's DGX Spark hardware. Perplexity, valued at more than $30 billion, has been drifting this way for a while: it integrated Zhipu's open-weight GLM 5.2 in July and shipped a Hybrid Compute feature this month that splits a job between the device and the cloud.
The important part is not the model choice, it's the routing. Perplexity says work completed locally doesn't burn its cloud credits, which turns an agent that runs for hours into a fixed cost instead of a metered one. Harder steps — fresh web data, connected apps, stronger reasoning — still escalate to the cloud tier, where Perplexity offers access to more than 15 frontier models. Perplexity keeps the interface, the connectors and the escalation path even when it supplies neither the silicon nor the model. That is a real business, but it is an orchestration business, and orchestration businesses are easier to copy than model businesses — every startup with a good runtime can swap in whichever open weights win the week.
Arm is now shipping a CPU line called "AGI CPU," plus Neoverse CSS N4, a configurable compute subsystem with up to 128 cores per die. CSS N4 supports LPDDR6 memory and PCIe Gen 7 and claims up to twice the performance, 1.25x the performance per watt and 1.75x the memory bandwidth of the previous generation. Arm's pitch is that agents move the bottleneck off the accelerator: retrieving data, calling tools and coordinating other chips is CPU work, and its reference design packs 8,160 cores into a standard air-cooled rack.
The customer list is the actual news. Arm says OpenAI, Meta, Cloudflare, Oracle, SAP, Lenovo and Supermicro are building around the AGI CPU, with Meta as lead partner, and ByteDance's Volcano Engine preparing what Arm calls the first agentic sandboxes on the chip. Google is already running agent sandboxes on its Arm-based Axion processors and Microsoft on Cobalt 200 — the pattern Arm is selling is that the whole industry standardizes on its instruction set whether they build custom silicon or buy off the shelf.
SentinelOne expanded its Wayfinder frontier AI services to run OpenAI's Daybreak models. Daybreak is the tiered program OpenAI built to put its strongest cyber-capable models into vetted defenders' hands rather than releasing them broadly; SentinelOne is now one of those hands, wiring them into managed detection and response. It is the third big security vendor to take this route this month, and it confirms the shape OpenAI chose: capability distributed through trusted intermediaries, with access control doing the work that a release decision used to do.
What to watch: whether Perplexity's cloud escalation path holds its margins as open-weight local models keep improving — and whether the defenders getting Daybreak access can show results fast enough to justify keeping it gated.
Would you run a work agent on your own hardware if it meant giving up frontier-model quality on the hard steps? Tell us in the comments.
