When NOT to run local AI.

1 · Frontier reasoning

If your work depends on the current frontier — Claude Opus 4.7, GPT-5.5 (April 23 2026), Gemini 3 Pro Thinking — local won’t catch up. Even a 5090 running Qwen 3.5 35B-A3B is roughly GPT-4- tier on mixed reasoning tasks: great for a lot of work, not great for the problems where you actually need the frontier. People who try to solve hard PhD-level problems locally eventually buy the Claude Max plan and stop pretending.

Cloud wins: complex multi-step reasoning, novel proofs, legal or medical synthesis where hallucination costs money.

2 · Anything multi-modal video

Local video understanding is still a research project. The cloud-native offerings — Gemini 3 Pro’s video input, Claude’s screenshare, OpenAI’s computer-use — work today at quality that no 24-GB consumer card touches. If your workflow includes “watch this 40-minute meeting recording and summarize”, cloud is the answer.

Cloud wins: video transcription + understanding, screen understanding, computer-use agents, long-form audio analysis at quality.

3 · Context beyond ~200K tokens

Qwen 3.5 27B natively handles 262K context and can be pushed to ~1M with YaRN. On paper that matches Claude and beats GPT-5. In practice, prefill latency on a Mac M5 Max at 147K context drops to ~666 PP tok/s and takes over a minute for first token. That’s fine for batch processing; it’s unusable for interactive work.

Cloud wins: interactive coding against a 300K-token codebase, long-document review where you want chat-like turnaround, any workflow where “first token” needs to be sub-5-seconds on a long prompt.

4 · Current-events / recall-heavy tasks

Local model weights are a frozen snapshot of the training cutoff. Qwen 3.6-27B (April 22 2026) has no knowledge of anything that happened after its training window, no matter how good its reasoning is. Cloud models have web-browse, RAG against fresh indices, and grounded search. If you need current information, local is not competitive on this axis.

Cloud wins: news, live pricing, research on emerging topics, fact-checking against the current internet.

5 · One-off heavy tasks

If you use AI lightly — a few hundred thousand tokens per month — the ROI on local hardware is decades, not months. The cost calculator on the homepage puts numbers on this: a $3,000 rig paying back a $20/month Plus plan takes ~12 years. Local only pays for itself if you use it heavily, have a privacy hard-requirement, or want the learning experience.

Cloud wins: light daily use, sporadic heavy tasks, anyone who doesn’t want to babysit a runner.

6 · When you’re fighting the ecosystem

Two honest cases where the community evangelism leads people astray: AMD ROCm in 2026 still takes 5–10 hours of driver setup, and Intel Arc with IPEX-LLM archived as of January 28, 2026 means that card’s software stack is in limbo. If you picked AMD or Intel because an enthusiast on Reddit said they were “fine” — they are, eventually. You will spend a weekend getting there. On NVIDIA or a Mac you were up in 15 minutes.

Cloud wins: when the friction of running local competes with the actual work you need to do.

So when DO you run local

Everything else. Privacy-bound workflows. High-volume coding or chat at zero marginal cost. Experimentation + tinkering. Offline reliability when travelling. Integration with tools that never leave your machine. Workflows where cost-per-token × tokens-per-day crosses the break-even line.

The answer is almost never “local only” or “cloud only.” It’s a hybrid — cloud for the hard 10% of tasks, local for the 90% where a well-chosen 35B MoE answers just as well for free. Use the planner to figure out which local setup covers your 90%.

Next step