text model · Sarvam · macOS
Can I run Sarvam-M 24B on Apple M3 Ultra (256GB)?
Yes. Sarvam-M 24B runs on Apple M3 Ultra (256GB) at Q4_K_M (~16.3 GB of ~192 GB usable).
Runs at Q4_K_M using ~16.3 GB of ~192 GB usable. You have room for FP16 for higher quality.
That figure is at a 4k context and moves about ±15% as context length changes. Apple M3 Ultra (256GB) leaves ~175.7 GB of headroom, room to step up to FP16 for higher quality.
- Q4_K_M needed
- ~16.3 GB
- Usable on device
- ~192 GB
- Device memory
- 256 GB
- Best quant
- Q4_K_M
Which quant fits
Running cost · estimate
- Power draw
- ~270 W
- Electricity / 1M tokens
- ~$0.24
- Pays for itself after
- ~15,381M tok
At ~$0.15/kWh and the estimated ~46 tok/s, a million generated tokens costs about $0.24 in electricity. Versus a hosted API at ~$0.5 per million tokens, the ~$3,999 Apple M3 Ultra (256GB) pays for itself after roughly 15,381 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
Pick your tool. All 2 load the same Q4_K_M weights.
llama-cli -hf lmstudio-community/sarvam-m-GGUF:Q4_K_M lms get lmstudio-community/sarvam-m-GGUF How to run it
On macOS use LM Studio (Polished GUI, ships MLX on Apple Silicon, one-click model downloads.).
- Parameters
- 24B
- Q4_K_M size
- 14.3 GB
- Q8_0 size
- 25.1 GB
- Context
- 32k
- Memory
- 256 GB unified
- Usable for weights
- ~192 GB
- Power draw
- ~270 W
- Best runtime
- MLX direct / Ollama (MLX backend)
You could also run
Run Sarvam-M 24B on other hardware
FAQ
Can Apple M3 Ultra (256GB) run Sarvam-M 24B?
Yes. Sarvam-M 24B runs on Apple M3 Ultra (256GB) at Q4_K_M (~16.3 GB of ~192 GB usable).
How much memory does Sarvam-M 24B need?
Apple M3 Ultra (256GB) has room to spare. At Q4_K_M the weights are ~14.3 GB; with KV cache and runtime overhead, budget ~16.3 GB at a 4k context.
What is the best tool to run Sarvam-M 24B 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/sarvam-m-24b/apple-m3-ultra-256gb) 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.