Memory at each precision
| Version | Weights | With context | Smallest Mac |
|---|---|---|---|
| native 4.25-bit | 13.8 GB | 15.3 GB | 24 GB |
| 3-bit lower quality | 9.2 GB | 10.7 GB | 16 GB |
| 2-bit lower quality | 6.6 GB | 8.1 GB | 16 GB |
Ships in a native low-bit format: about 14 GB of weights.
On every Apple chip
The smallest memory size of each chip that runs gpt-oss-20b, the best version that fits it, and the writing speed you can expect.
| Chip | Needs | Version | Tokens/s |
|---|---|---|---|
| M6 | 24 GB | native 4.25-bit Runs well | 40–72 |
| M5 Ultra | 96 GB | native 4.25-bit Runs well | 284–505 |
| M5 Max | 36 GB | native 4.25-bit Runs well | 109–194 |
| M5 Pro | 24 GB | native 4.25-bit Runs well | 73–129 |
| M5 | 24 GB | native 4.25-bit Runs well | 36–64 |
| M4 Max | 36 GB | native 4.25-bit Runs well | 97–173 |
| M4 Pro | 24 GB | native 4.25-bit Runs well | 65–115 |
| M4 | 24 GB | native 4.25-bit Runs well | 28–51 |
| M3 Ultra | 96 GB | native 4.25-bit Runs well | 194–345 |
| M3 Max | 36 GB | native 4.25-bit Runs well | 71–126 |
| M3 Pro | 36 GB | native 4.25-bit Runs well | 36–63 |
| M3 | 24 GB | native 4.25-bit Runs well | 24–42 |
| M2 Ultra | 64 GB | native 4.25-bit Runs well | 189–337 |
| M2 Max | 32 GB | native 4.25-bit Runs well | 95–168 |
| M2 Pro | 32 GB | native 4.25-bit Runs well | 47–84 |
| M2 | 24 GB | native 4.25-bit Runs well | 24–42 |
| M1 Ultra | 64 GB | native 4.25-bit Runs well | 189–337 |
| M1 Max | 32 GB | native 4.25-bit Runs well | 95–168 |
| M1 Pro | 32 GB | native 4.25-bit Runs well | 47–84 |
| M1 | 16 GB | 3-bit Runs well | 19–34 |
How to run it
- The easy way: download MLX Harness, free and open source. Search for gpt-oss-20b, download the version that fits your Mac, and chat. It also serves the model to your other apps over an OpenAI-compatible API.Download MLX HarnessFree · Apple Silicon · macOS 14.5 or later
- In Terminal with Apple’s MLX:
pip install mlx-lmmlx_lm.chat --model mlx-community/gpt-oss-20b-MXFP4-Q8Build: mlx-community/gpt-oss-20b-MXFP4-Q8 on Hugging Face. Other precisions sit next to it.
Step by step: download and run LLMs on a Mac · how to use MLX Harness · MLX vs Ollama vs llama.cpp
About gpt-oss-20b
| Made by | OpenAI |
|---|---|
| Parameters | 21B total · 3.6B active (mixture of experts) |
| Context | 128K tokens |
| Licence | Apache-2.0 |
| Released | 2025-08 |
| Input | Text |
| Intelligence | 9 on the Artificial Analysis Intelligence Index |
| Weights | openai/gpt-oss-20b |
Questions
- How much memory does gpt-oss-20b need?
- About 13.8 GB of weights at native 4.25-bit, plus room for the context: 15.3 GB in all. On a Mac that means 24 GB of unified memory or more, since macOS lets the GPU use about 75% of it by default.
- How fast is gpt-oss-20b on a Mac?
- Writing speed depends on memory bandwidth. The fastest Mac for it here is an M5 Ultra with 96 GB at about 284–505 tokens a second (native 4.25-bit). The smallest Mac that runs it, with 24 GB, manages about 73–129 on an M5 Pro.
- Why does gpt-oss-20b run faster than its size suggests?
- It is a mixture-of-experts model: it holds 21B parameters in memory but reads only 3.6B for each token, so it writes at the speed of a much smaller model.
Similar models
For your model card or README
Shows the smallest Mac that runs gpt-oss-20b well and links here.
[](https://mlx.app/run/gpt-oss-20b)How it’s worked out
Memory: the model’s parameters at 8, 6, 4, 3 or 2 bits per weight plus MLX’s group scales (half a bit per weight), or its native format, with room for the context on top. By default macOS lets the GPU use about 75% of unified memory. Speed: memory bandwidth divided by the bytes read per token is the ceiling for one conversation; in llama.cpp’s public Apple Silicon benchmarks real runs reach about 45% to 80% of it, which is the range shown. These are estimates, not measurements: measure your own Mac on MLX Bench. Model figures from each official model card, chip figures from Apple’s Tech Specs. Checked 2026-10-19. Same data as mlx.app/compare.