All capabilities Local and private

Run a capable model on your own machine

Open-weight models on your own hardware, where the constraint is memory arithmetic rather than a price per token.

DocsDocumented in partTestNot tested by benchr Medium · Local Docs checked September 3, 2026
Local runtimesQwen3.6-27BLlama 4 ScoutPhi-4

Overview

  • Runs entirely on your machine, so no prompt or document leaves it.
  • Removes per-token cost, replacing it with a fixed hardware ceiling.
  • Keeps working offline, which matters more than it sounds for regulated and field work.

How it works

  1. 01Start from memory, not from benchmarks. benchr's planner estimates four-bit weights as total parameters multiplied by 0.5 GB, and that figure excludes runtime buffers, context cache and the operating system.
  2. 02Check the fit label before downloading anything: comfortable, tight, or not recommended.
  3. 03Treat an advertised extended context as expensive, not free - it can need runtime-specific scaling and substantially more cache memory.
  4. 04Compare against a hosted small model honestly. Local wins on privacy and marginal cost, not usually on quality per watt.

Limits

  • The memory figures are deployment-planning estimates, not throughput measurements or guarantees.
  • Where a provider does not publish quantised memory, the estimate is arithmetic, and benchr labels it as such.
  • Extended context is not free memory - the planner explicitly refuses to treat it that way.
  • This record is checked against benchr's own local deployment index rather than a single vendor page, and that index carries its own version date.

Verification

benchr test runs

benchr has not run this capability. No test has been published.

benchr test runs →

Evidence

Not stated by the source

  • Tokens per second on any specific machine
  • Quality parity with a hosted frontier model on your workload
Source-check history
  • September 3, 2026DocsDocumented in partAdded an official local-runtime source. The record previously rested on benchr's own deployment index, which is not a vendor source and cannot support a documented status on its own.
  • September 1, 2026DocsDocumented in partFirst logged as documented in part. The evidence here is benchr's own sourced local index rather than a single vendor page, and every figure is a planning estimate, so the record does not claim full official documentation.

Changes

Provider changes affecting this