Perplexity Portable Computer: Local Agents on NVIDIA DGX Spark

On August 25, 2026, Perplexity posted a demo video of a fully local Perplexity Computer running on NVIDIA DGX Spark — orchestrator LLM, subagent LLM, and agent harness on the box, with cloud routing optional and user-gated. The thread drew roughly 473K views on X, and the reaction split predictably: local-AI builders saw validation; cloud-Computer subscribers asked whether their $200/month Max plan now includes a $4,679 hardware bill.

It does not. Local Computer is a Pro/Max feature tied to DGX Spark ownership — a productized version of the personal local AI stack explainx.ai has been documenting all year, wrapped in Perplexity's Computer harness instead of a DIY terminal workflow.

TL;DR: Local vs cloud Computer

What Perplexity actually demoed

The video shows Perplexity Computer's full agent stack — not just a chat model in a browser tab — running locally:

1. Orchestrator LLM — plans multi-step tasks, delegates to subagents
2. Subagent LLM — executes scoped work units under orchestrator direction
3. Agent harness — tool calls, file handling, the same Computer UX cloud users know from Projects and Brain memory

Local models shipped or named in the demo:

• PPLX 27B (post-trained) — default local orchestrator
• Qwen 3.8 27B — available local option; explainx.ai benchmarked Qwen 3.8 extensively on DGX Spark in our Qwen 3.8 27B guide
• Nemotron 3.5 Lightning — coming soon; see Nemotron 3.5 on DGX Spark for NVIDIA's local serving recipe

Frontier reasoning gate: when a task exceeds local model capability, Perplexity routes to cloud — but only after explicit user approval, with PII-flagged content blocked from upload and text guidance only (not full document exfiltration). Sensitive files remain on the Spark.

That is the product thesis in one sentence: Computer's agent UX, with a local trust boundary.

Why DGX Spark specifically

Perplexity did not pick random gaming hardware. The NVIDIA DGX Spark is a 128GB unified-memory Grace Blackwell box priced around $4,679 — the same machine explainx.ai recommended for running 200B-class models locally with native CUDA support.

For Perplexity's stack:

• 27B orchestrator + 27B subagent fits comfortably in unified memory with room for KV cache and tool context
• CUDA ecosystem means Perplexity can reuse NVIDIA's serving optimizations (including future Nemotron DSpark paths)
• Desktop form factor matches the "portable Computer" branding — lug it between office and home, not rack-mount a cluster

Compare to Liquid AI's LFM2.5 2.6B on-device agents for the opposite end of the spectrum: tiny models on phones vs Perplexity's full Computer harness on a desktop superchip.

How this fits Perplexity's 2026 strategy

Perplexity spent the summer optimizing cloud Computer economics — GLM 5.2 orchestrator at 0.34× Opus cost, Grok orchestrator swaps, and pass-through Sol pricing after OpenAI's 20% cut. Local Computer is the sovereignty branch of the same tree: same harness, different inference location.

What people are asking

Is this available to download today? Perplexity showed a working demo; general availability timing was not confirmed in the August 25 video. Treat it as a directional product preview, not something you can apt-install tonight.

Teknium's "This app looks familiar" reaction? Open-source builder Teknium (@Teknium) noted visual similarity to existing local agent UIs — fair observation given Perplexity's harness converges on patterns the open-source agent ecosystem (OpenCode, Hermes Agent, etc.) pioneered. Similarity does not negate value; integration and model routing still matter.

Can I use a 4090 instead of DGX Spark? Perplexity specified DGX Spark in the demo. A 24GB 4090 can run quantized 27B models, but dual-model orchestrator+subagent concurrency is tighter — see Mac vs dedicated GPU economics for the general tradeoff math.

Does local Computer replace WANDR / Brain / Search as Code? Cloud-only features (overnight Brain graphs, hosted Search as Code) may not have local parity on day one. Expect a subset of cloud Computer capabilities initially.

Honest limitations

• Hardware tax is real — $4,679 plus Pro/Max subscription before you run one local task. Cloud Computer remains cheaper for intermittent use.
• 27B is not frontier — user-gated cloud routing exists because local models will fail hard reasoning tasks Qwen/GLM/Sol handle in cloud mode.
• Agent security ≠ local security — keeping docs on-device helps confidentiality; it does not stop a misconfigured agent from deleting local files or calling bad tools.
• Nemotron "coming soon" is not shipped — plan around PPLX 27B and Qwen 3.8 27B only for now.
• Not a multi-user server — one Spark, one operator pattern; teams need separate boxes or cloud Computer with Projects.

Build your own vs wait for Perplexity

If you already own a DGX Spark, you can approximate pieces today:

# Illustrative stack — not Perplexity's exact bundle
ollama pull qwen3.8:27b
# Pair with OpenCode or a local agent harness — see explainx.ai guides

For the integrated Perplexity UX, waiting for their DGX Spark bundle may be simpler than replicating orchestrator+subagent routing yourself. For maximum control, build a personal local AI system and swap models freely.

Related on explainx.ai

• NVIDIA DGX Spark: best local LLM setup
• Qwen 3.8 27B on DGX Spark — benchmarks
• NVIDIA Nemotron 3.5 Lightning — DGX Spark recipe
• Perplexity Computer GLM 5.2 orchestrator — cloud contrast
• Perplexity Sol price pass-through after OpenAI cut
• Build a personal local AI system
• Liquid AI LFM2.5 — on-device agents at the other extreme
• Run open-source models locally in OpenCode

Official: Perplexity demo video on X · Perplexity Computer

Hardware requirements and model availability reflect Perplexity's August 25, 2026 demo. Local Computer GA timing and feature parity with cloud Computer are unconfirmed as of this writing.