Perplexity's 'Portable Computer' Runs on Alibaba's Qwen3.8-27B, Not Claude or GPT
The AMW Read
Confirms and extends Perplexity's known shift toward open-weight models (GLM 5.2, Hybrid Compute) by revealing Qwen3.8-27B powers its already-launched Portable Computer agent.
Named counterparties: Alibaba Group
Perplexity's 'Portable Computer' Runs on Alibaba's Qwen3.8-27B, Not Claude or GPT
Perplexity's on-device AI agent Portable Computer — first introduced in partnership with Nvidia on DGX Spark hardware — runs on Alibaba's open-source Qwen3.8-27B model rather than Anthropic Claude or OpenAI GPT, according to a September 8 PingWest report citing SiliconAngle. Perplexity built a custom "PPLX 27B" optimization layer on top of Qwen3.8-27B to handle local file processing, data analysis and coding tasks on-device.
The disclosure extends a pattern already visible in Perplexity's product roadmap: after integrating Zhipu AI's open-weight GLM 5.2 in July and rolling out the on-device/cloud-split Hybrid Compute feature in early September, the company is increasingly routing its agent stack to open-weight models rather than frontier closed labs, even as Nvidia has reportedly been in talks to invest in Perplexity at a valuation above $30 billion. Qwen's rapid string of releases (Qwen3.8-Max, Qwen3.8-27B, Qwen3.8-Flash) has been adopted well beyond Perplexity — the article cites Airbnb and Pinterest as production users and Reuters as building on it — pointing to open-weight Chinese models becoming a viable default layer for Western AI products, not just a low-cost fallback.
For builders, this signals falling switching costs between closed frontier APIs and open-weight local models on capable hardware like DGX Spark, especially for latency-sensitive, no-token-cost agent workloads. For investors, Perplexity leaning on a third-party open-weight model rather than its own or a partner's proprietary model is worth watching against its reported $30 billion-plus valuation talks with Nvidia — differentiation looks like it's shifting toward orchestration and local deployment rather than model ownership.
