Memory at each precision
| Version | Weights | With context | Smallest Mac |
|---|---|---|---|
| native 4.25-bit | 1.6 TB | 1.7 TB | more than any Mac |
| 3-bit lower quality | 1.2 TB | 1.3 TB | more than any Mac |
| 2-bit lower quality | 875 GB | 945 GB | more than any Mac |
Ships in a native low-bit format: about 1561 GB of weights.
On every Apple chip
The smallest memory size of each chip that runs Kimi K3, the best version that fits it, and the writing speed you can expect.
| Chip | Needs | Version | Tokens/s |
|---|---|---|---|
| M6 | no size fits | Too big | – |
| M5 Ultra | no size fits | Too big | – |
| M5 Max | no size fits | Too big | – |
| M5 Pro | no size fits | Too big | – |
| M5 | no size fits | Too big | – |
| M4 Max | no size fits | Too big | – |
| M4 Pro | no size fits | Too big | – |
| M4 | no size fits | Too big | – |
| M3 Ultra | no size fits | Too big | – |
| M3 Max | no size fits | Too big | – |
| M3 Pro | no size fits | Too big | – |
| M3 | no size fits | Too big | – |
| M2 Ultra | no size fits | Too big | – |
| M2 Max | no size fits | Too big | – |
| M2 Pro | no size fits | Too big | – |
| M2 | no size fits | Too big | – |
| M1 Ultra | no size fits | Too big | – |
| M1 Max | no size fits | Too big | – |
| M1 Pro | no size fits | Too big | – |
| M1 | no size fits | Too big | – |
How to run it
No single Mac holds it. Smaller models below run well, and the guide to Mac clusters covers spreading a model over several Macs.
Step by step: download and run LLMs on a Mac · how to use MLX Harness · MLX vs Ollama vs llama.cpp
About Kimi K3
| Made by | Moonshot AI |
|---|---|
| Parameters | 2,800B total · 104B active (mixture of experts) |
| Context | 1M tokens |
| Licence | Kimi K3 License (custom) |
| Released | 2026-06 |
| Input | Text and images |
| Intelligence | 43.6 on the Artificial Analysis Intelligence Index |
| Weights | moonshotai/Kimi-K3 |
Questions
- How much memory does Kimi K3 need?
- About 1.6 TB of weights at native 4.25-bit, plus room for the context: 1.7 TB in all. That is more than any Mac has today.
- Can any Mac run Kimi K3?
- Not on its own: even at 2 bits it needs more memory than the largest Mac. Use a smaller model, or spread it over several Macs with a cluster runtime.
- Why does Kimi K3 run faster than its size suggests?
- It is a mixture-of-experts model: it holds 2,800B parameters in memory but reads only 104B 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 Kimi K3 well and links here.
[](https://mlx.app/run/kimi-k3)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.