text model · Mistral · macOS
Can I run Magistral Small on Apple M4 (24GB)?
Yes. Magistral Small runs on Apple M4 (24GB) at Q4_K_M (~16 GB of ~16 GB usable).
Fits at Q4_K_M (~16 GB of ~16 GB usable) but with little headroom. Close other apps, or drop to a 2k context to free about a gigabyte.
That figure is at a 4k context and moves about ±15% as context length changes.
- Q4_K_M needed
- ~16 GB
- Usable on device
- ~16 GB
- Device memory
- 24 GB
- Best quant
- Q4_K_M
Which quant fits
Running cost · estimate
- Power draw
- ~65 W
- Electricity / 1M tokens
- ~$0.39
- Pays for itself after
- ~11,809M tok
At ~$0.15/kWh and the estimated ~7 tok/s, a million generated tokens costs about $0.39 in electricity. Versus a hosted API at ~$0.5 per million tokens, the ~$1,299 Apple M4 (24GB) pays for itself after roughly 11,809 million tokens, so local hardware is mostly a fixed cost, not a per-token one. TDP is the peak draw, so this is an upper bound. Assumptions.
Run it
Loads the Q4_K_M weights.
ollama run magistral:24b - Parameters
- 24B
- Q4_K_M size
- 14 GB
- Q8_0 size
- 25 GB
- Context
- 40k
- Ollama tag
- magistral:24b
- Memory
- 24 GB unified
- Usable for weights
- ~16 GB
- Power draw
- ~65 W
- Best runtime
- Ollama (MLX backend, preview) / MLX direct
You could also run
Run Magistral Small on other hardware
FAQ
Can Apple M4 (24GB) run Magistral Small?
Yes. Magistral Small runs on Apple M4 (24GB) at Q4_K_M (~16 GB of ~16 GB usable).
How much memory does Magistral Small need?
It is a tight fit on Apple M4 (24GB). At Q4_K_M the weights are ~14 GB; with KV cache and runtime overhead, budget ~16 GB at a 4k context.
What is the best tool to run Magistral Small on macOS?
LM Studio for a simple setup; mlx-lm for the most speed. vLLM is NOT a Mac tool, it is a CUDA/Linux serving engine. Unified memory is not a fixed VRAM slice; ~70% is usable for weights.
Embed this
A live badge for your model card or README, updated as the data is. [](https://localmodel.run/can-i-run/magistral-small/apple-m4-24gb) Sources
Weights are measured from GGUF files; KV cache and overhead are computed, so totals can vary ~15% with context and runtime. Any tok/s is a bandwidth estimate. See methodology.