Overview
The gpt-oss Mac requirements are about 15.96 GB of unified memory for gpt-oss-20b and about 67.55 GB for gpt-oss-120b at a 4096-token context, using this site’s formula plus the actual file sizes on Hugging Face. So the 20B model is a 24 GB Mac model (16 GB is on the edge and unconfirmed) and the 120B model needs a 128 GB Mac, not 64 GB. Both are Apache-2.0 models from OpenAI with MXFP4-quantized mixture-of-experts weights. Method: methodology. Speed is unknown on this site; see measure-tokens-per-second-mac.
What do the gpt-oss model cards say?
The cards give parameter counts, active parameters and the quantization, and one memory statement for each size.
| Item | gpt-oss-20b | gpt-oss-120b |
|---|---|---|
| Parameters (card) | 21B, 3.6B active | 117B, 5.1B active |
| Parameters (Hugging Face safetensors count) | 20,914,757,184 | 116,829,156,672 |
| Layers / KV heads / head dim (config.json) | 24 / 8 / 64 | 36 / 8 / 64 |
| Max positions (config.json) | 131,072 | 131,072 |
| Card memory statement (vendor claim) | “run within 16GB of memory" | "fit into a single 80GB GPU” |
| License | Apache 2.0 | Apache 2.0 |
The cards say the MoE weights were post-trained with MXFP4 quantization. In a mixture-of-experts model all experts must be held in memory, so the 3.6B and 5.1B active figures do not shrink the memory need. The weights below use the total file size.
How much RAM does gpt-oss-20b need on a Mac?
At 4096 context the total is 15.962 GB, and at the full 131,072 context it is 22.203 GB.
The formulas are from methodology. Weights are taken as the published file size rather than params × bits / 8, because the MXFP4 checkpoint keeps attention, embeddings and the output layer in BF16, so no single bit width applies (the modules_to_not_convert list in config.json). The three safetensors files in the repo add to 13.761 GB (4,792,272,488 + 4,798,702,184 + 4,170,342,232 bytes). Ollama’s page lists the gpt-oss:20b tag at 14GB with a 128K context.
kv_GB = 2 × layers × kv_heads × head_dim × context × 2 / 1e9, with 2 bytes per value assumed. That gives 2 × 24 × 8 × 64 × 2 = 49,152 bytes per token.
| Context | KV cache | Weights 13.761 + KV + 2 GB overhead |
|---|---|---|
| 4,096 | 0.201 GB | 15.962 GB |
| 32,768 | 1.611 GB | 17.372 GB |
| 131,072 | 6.442 GB | 22.203 GB |
The KV figures count every layer as full attention. config.json marks every second layer sliding_attention with a 128-token window, so the real cache is lower. By how much is unknown here, which makes these totals an upper bound. The 2 GB overhead is this site’s assumption, not a measurement.
Can I run gpt-oss-20b on a 16GB Mac mini M4?
Possibly, but not with margin: 15.962 GB at 4096 context leaves 0.038 GB under 16 GB before the macOS reserve, which is unknown. Apple’s Mac mini (2024) page lists the M4 with 16GB unified memory, configurable to 24GB or 32GB, and 120GB/s memory bandwidth. The card’s “within 16GB” is OpenAI’s claim, and the card does not state a context length or overhead behind it. best-llm-for-16gb-mac rates this model “Tight” using a 4.0-bit estimate of 10.5 GB for the weights. The real file is 13.761 GB, which is why this page’s total is higher. On a 16 GB Mac, keep other apps closed and expect the load to possibly fail. The GPU-side limit is covered on mac-gpu-memory-limit.
How much RAM does gpt-oss-120b need on a Mac?
At 4096 context the total is 67.551 GB, and at 131,072 it is 76.913 GB, so 64 GB is not enough.
The 15 safetensors files in the repo add to 65.249 GB. Ollama lists the gpt-oss:120b tag at 65GB. KV per token is 2 × 36 × 8 × 64 × 2 = 73,728 bytes, again counting every layer as full attention.
| Context | KV cache | Weights 65.249 + KV + 2 GB overhead |
|---|---|---|
| 4,096 | 0.302 GB | 67.551 GB |
| 32,768 | 2.416 GB | 69.665 GB |
| 131,072 | 9.664 GB | 76.913 GB |
The weights alone (65.249 GB) already exceed 64 GB. The card’s “single 80GB GPU” is a data-centre statement and is not a Mac figure.
Which Mac do I need for gpt-oss?
Compare the totals with installed unified memory, before the unknown macOS reserve.
| Unified memory | Where Apple lists it | gpt-oss-20b (15.962 to 22.203 GB) | gpt-oss-120b (67.551 to 76.913 GB) |
|---|---|---|---|
| 16 GB | Mac mini M4 | Edge at 4096; no at longer contexts | No |
| 24 GB | Mac mini M4, MacBook Pro M5 Pro base | Yes, including 131,072 (22.203 GB), little room | No |
| 32 GB | Mac mini M4, MacBook Pro M5 | Yes | No |
| 64 GB | MacBook Pro M5 Pro or M5 Max, Mac mini M5 Pro | Yes | No |
| 96 GB | Mac Studio M5 Ultra base | Yes | Yes on paper, even at 131,072 |
| 128 GB | MacBook Pro M5 Max, Mac Studio M5 Max | Yes | Yes on paper, even at 131,072 |
“Yes” means the total is under installed memory. It is not a load test. Whether a 120B model loads on a given Mac depends on the unknown macOS reserve. Picks by tier: best-llm-for-24gb-mac, best-llm-for-32gb-mac, best-llm-for-128gb-mac.
How do I run gpt-oss-120b with Ollama on a Mac?
Install Ollama (see ollama-on-mac) and run the command from Ollama’s gpt-oss page: ollama run gpt-oss:120b. For the 20B model use ollama run gpt-oss:20b. The 120B tag is a 65GB download. The Hugging Face 20B card also gives lms get openai/gpt-oss-20b for LM Studio. Whether a given Mac runs 120B well is unknown; no speed figure is published here.
Why do gpt-oss memory minimums disagree?
Sources disagree because they set different context lengths, overhead and units. 15.962 GB at 4096 context and 22.203 GB at the full 131,072 context are both correct for the same file. A “16GB” claim is the vendor’s statement for the 20B model without stated conditions, and the model card does not give an 8 GB figure. Show the context and overhead with any number, as above.
Questions
Can I run gpt-oss 120b on a Mac?
Yes on a Mac with 96 GB or more of unified memory on paper, no on 64 GB or less. The weights are 65.249 GB and the total at 4096 context is 67.551 GB. The macOS reserve is unknown, so a 96 GB or 128 GB Mac is unconfirmed until you load it.
Can I run gpt-oss 20b on a Mac?
Yes on a 24 GB or larger Mac. Its total is 15.962 GB at 4096 context and 22.203 GB at 131,072. A 16 GB Mac is on the edge and unconfirmed.
Can gpt-oss 20b run on a Mac mini M4 with 16GB?
It is borderline. The total is 15.962 GB against 16GB installed, with an unknown macOS reserve, so it may not load. The 24GB Mac mini M4 configuration is clearly enough.
How much RAM does gpt-oss 120b need?
The weights are 65.249 GB and the total is 67.551 GB at 4096 context, up to 76.913 GB at 131,072, using a 2 GB overhead assumption and an all-full-attention KV upper bound.
What are the gpt-oss Mac requirements?
20B: at least 24 GB for comfortable use. 120B: 96 GB or more, so in practice 128 GB. Both are computed from methodology and the Hugging Face file listing.
How do I run gpt-oss-120b with Ollama on a Mac?
Run ollama run gpt-oss:120b after installing Ollama, on a Mac with enough memory per the table above. The tag is listed at 65GB.
Is gpt-oss fast on a Mac?
Unknown. This site publishes no speed figures; measure your own Mac with measure-tokens-per-second-mac.
Sources read on 2026-10-13
- gpt-oss-20b and gpt-oss-120b cards: parameters, active parameters, MXFP4, memory statements, license, run commands.
- 20b config.json and 120b config.json: layers, KV heads, head dim, sliding window, max positions, unquantized modules.
- Hugging Face model API for 20b and 120b: safetensors parameter counts and file sizes.
- Ollama gpt-oss: tag sizes, 128K context, run commands.
- Mac mini (2024) specs, Mac mini, MacBook Pro and Mac Studio spec pages: memory options and bandwidth.
- methodology: weight, KV and overhead formulas.