Model Picks That Fit Your Machine

June 28, 2026

The model picker now knows how much memory your device has and recommends the right local model for it — one strong primary per RAM tier, instead of a generic list you have to size yourself. If a model can't fit, it won't be recommended.

What you can do

• See your device's memory in the picker and get local-model recommendations curated to it. The picker reads your available AI memory and sorts you into a band — under 16GB, 16–32, 32–64, 64–128, or 128GB and up.
• Get one strong local primary per RAM tier instead of a flat list. Each recommended local model carries the memory floor it needs, so the picker only surfaces models that actually run well on your machine.
• Trust cloud recommendations on a server. On a pure cloud setup with no local runtime, the picker skips the memory probe entirely and shows cloud picks only — no misleading local suggestions.
• Run bigger inputs on GPT-OSS 120B. Its context window grew from 32K to 131K tokens, so it can take much larger documents and codebases in a single pass.

Where this shows up

• You open the picker on a 32GB laptop. Instead of seeing a 70B model that would thrash your memory, you get the strongest local model that genuinely fits — and a clear note of how much memory you have to work with.
• You move to a 128GB workstation. The picker now recommends a larger, more capable local primary, because it can see the headroom.
• You're running Alfrada on a cloud VPS with no local models. The picker shows cloud options cleanly, without a "recommended local" section that wouldn't apply.

Try it

• "Open the model picker and pick the best local model my machine can run."
• "Summarize this 100K-token document with GPT-OSS 120B."

Heads up

• Memory-based recommendations only appear when a local runtime is present. On cloud-only setups, that part of the picker stays hidden by design.