OpenAI · Apache 2.0
gpt-oss
OpenAI’s open models for reasoning and tool use. The 20B fits a 16 GB card and is fast.
How much memory it needs
The model file, plus memory for the conversation, plus about 1 GB for the app. Longer conversations need more.
| Size | Version | File | At 8K | At 32K | At 128K |
|---|---|---|---|---|---|
| 21B (3.6B active) | MXFP4 | 11.6 GB | 12.8 GB | 13.4 GB | 15.9 GB |
| 117B (5.1B active) | MXFP4 | 62.8 GB | 64.1 GB | 65.0 GB | 68.6 GB |
Which computers run it
The best size for each machine at an 8K-token conversation. Speeds are estimates from memory bandwidth.
| Computer | Best version | Fit | Speed |
|---|---|---|---|
| Laptop or PC, no graphics card, 32 GB | 21B MXFP4 | Runs great | 15–30tok/sSmooth |
| RTX 4060 · 8 GB | 21B MXFP4 | Spills into system RAM | 25–45tok/sSmooth |
| RTX 3060 · 12 GB | 21B MXFP4 | Spills into system RAM | 45–85tok/sFast |
| RTX 5060 Ti · 16 GB | 21B MXFP4 | Runs well | 60–120tok/sFast |
| RTX 4090 · 24 GB | 21B MXFP4 | Runs great | 100–210tok/sFast |
| RTX 5090 · 32 GB | 21B MXFP4 | Runs great | 140–280tok/sFast |
| Mac with M4 · 16 GB | 21B MXFP4 | Fits with a tweak | 25–50tok/sSmooth |
| Mac with M4 Pro · 48 GB | 21B MXFP4 | Runs great | 40–85tok/sFast |
| Mac with M5 Max · 128 GB | 117B MXFP4 | Runs great | 50–110tok/sFast |
| Ryzen AI Max+ 395 · 128 GB | 117B MXFP4 | Runs great | 30–55tok/sFast |
| DGX Spark · 128 GB | 117B MXFP4 | Runs great | 30–60tok/sFast |
| RTX PRO 6000 · 96 GB | 117B MXFP4 | Runs great | 120–240tok/sFast |
| Mac Studio with M5 Ultra · 256 GB | 117B MXFP4 | Runs great | 50–110tok/sFast |
| Mac Studio with M3 Ultra · 512 GB | 117B MXFP4 | Runs great | 40–90tok/sFast |
Measured speeds
Real runs, not estimates. All measurements.
| Computer | Version | Prompt | Speed | Source |
|---|---|---|---|---|
| NVIDIA DGX Spark (GB10, 128 GB LPDDR5x) | MXFP4 | 512 | 58.72 tok/s | 2026-02-05 |
| AMD Ryzen AI Max+ 395 "Strix Halo" / Radeon 8060S (Framework Desktop, 128 GB LPDDR5x-8000) | MXFP4 | 512 | 33.7 tok/s | 2025-08 |
| Apple M3 Ultra (80-core GPU, 512 GB) | MLX 8-bit | 1K | 67.8 tok/s | 2026-08-06 |
| Apple M5 Max (18-core CPU, 40-core GPU, 14-inch MacBook Pro, 128 GB) | MXFP4 (MLX, 8-bit) | 4K | 87.9 tok/s | 2026-03-11 |
| NVIDIA GeForce RTX 3090 24 GB (used) | MXFP4 | 4K | 147.5 tok/s | 2026-03 |
| NVIDIA GeForce RTX 5060 Ti 16 GB | MXFP4 | 128K | 43.8 tok/s | 2026-03 |
| NVIDIA DGX Spark (GB10, 128 GB LPDDR5x) | MXFP4 | 512 | 86.23 tok/s | 2025-11-16 |
| Apple M4 Max (40-core GPU, MacBook Pro, 64 GB) | MXFP4 | 512 | 117.83 tok/s | 2025-11-16 |
What it is good and bad at
Good at
- Reasoning with adjustable effort
- Tool use and agents
- Fast: few parameters work per word
Not good at
- Text only
- Shorter 128K context than newer models
Start it
Install Ollama, then run one of these in a terminal. LM Studio users can search the same name in the app.
ollama run gpt-oss:120b