What we learned running Bonsai 27B on a 24 GB Mac

Updated 2026-07-15

Yes for moderate text work with one loaded runtime. LM Studio reported a 7.94 GiB loaded size at a 4,096 token context on our 24 GB M4 Pro Mac. The machine had enough room for the model, macOS, and the benchmark, but loading a second model server would have reduced the safety margin.

The choices that kept the run stable

  • Only one Bonsai 27B runtime was loaded.
  • The context limit was 4,096 tokens for the controlled speed test.
  • Other large applications stayed idle during measurement.
  • The existing model files were reused instead of copied.

Where 24 GB becomes tight

Long context adds cache memory. Vision work adds image processing costs. A second server would also load another model copy or runtime state.

A model that fits at 4K context is not proof that its full 262K context fits. Increase context in steps and watch memory pressure.

Questions people ask

Can you run two Bonsai 27B servers on 24 GB?

It may be possible in a narrow setup, but it leaves little room for context and the operating system. We did not use two servers for the controlled test.

Does the 8.49 GB file equal runtime memory?

No. File size and loaded memory are different measurements, and context adds more memory.

Sources

The benchmark numbers on this page describe one checkpoint, runtime, machine, and test shape. Reproduce the test on the hardware and workload you plan to use before making a product decision.