Want the details on one model? See the model guides, or compare with cloud GPU prices.
Try it live
Ask a question and watch a model answer in real time on our Mac Studio M4 Max 128GB in our datacenter. 50 free questions per hour.
Demo answers come straight from the model and may be wrong. Questions aren't saved. Want to ask follow-up questions? Open the full live chat →
Measured on our hardware
Real speeds from the live demo, not estimates. Each answer is timed on the machine shown, and this table updates as visitors use the demo.
| Model | Machine | Average speed | Best | Answers timed |
|---|---|---|---|---|
| lfm3:latest (nvfp4) try it | Mac Studio M4 Max 128GB | 209 tokens/sec | 210 | 2 |
| llama3.2:3b (3.2B Q4_K_M) try it | Mac Studio M4 Max 128GB | 166 tokens/sec | 172 | 10 |
| ministral-3:3b (3.8B Q4_K_M) try it | Mac Studio M4 Max 128GB | 150 tokens/sec | 153 | 9 |
| flint:40b (int4) try it | Mac Studio M4 Max 128GB | 146 tokens/sec | 151 | 6 |
| qwen3.5:4b (nvfp4) try it | Mac Studio M4 Max 128GB | 111 tokens/sec | 116 | 4 |
| gpt-oss:latest (20.9B MXFP4) try it | Mac Studio M4 Max 128GB | 98 tokens/sec | 99 | 13 |
| bonsai-2:latest (9.0B Q2_0) try it | Mac Studio M4 Max 128GB | 90 tokens/sec | 92 | 2 |
| glm-4.7-flash:latest (29.9B Q4_K_M) try it | Mac Studio M4 Max 128GB | 82 tokens/sec | 83 | 8 |
| qwen3-vl:latest (8.2B Q4_K_M) try it | Mac Studio M4 Max 128GB | 82 tokens/sec | 83 | 2 |
| deepseek-r1:latest (8.2B Q4_K_M) try it | Mac Studio M4 Max 128GB | 80 tokens/sec | 81 | 10 |
| gemma4:12b try it | Mac Studio M4 Max 128GB | 78 tokens/sec | 81 | 5 |
| ornith:latest (9.0B Q4_K_M) try it | Mac Studio M4 Max 128GB | 72 tokens/sec | 73 | 6 |
| qwen3.8:27b-mlx (nvfp4) try it | Mac Studio M4 Max 128GB | 50 tokens/sec | 60 | 3 |
See it running first
Book a demo on our Mac Studio M4 Max 128GB, or rent a Mac for a day and try your own model before you order. Every Mac also comes AI-ready, with Ollama, LM Studio, MLX, Open WebUI, Claude Code and more preinstalled free.
How we estimate
- Memory: model weights at the chosen precision, plus the context (KV) cache and runtime overhead. We leave about 10% of each Mac's memory for macOS.
- Speed: generating text on Apple Silicon is limited by memory bandwidth. We use Apple's published bandwidth for each chip, real-world efficiency of 45–65% (llama.cpp / MLX), and a small fixed cost per token. Mixture-of-experts (MoE) models only read their active experts, so they run much faster than their size suggests.
- Our hardware: every Mac we offer is the top configuration of its chip, with the most CPU and GPU cores.
- Clusters: the model's layers are split across Macs linked over Thunderbolt 5, so memory adds up while speed stays close to a single Mac's.
- Results are estimates. Prompt processing, batch size and software choice all affect real speed. Ask us if you want a benchmark on a specific machine.