Skip to content
localmodel.run

text model · Sarvam · macOS

S Can I run Sarvam-105B on Apple M3 Ultra (256GB)?

Compatibility verdict VRAM threshold engine
Yes, it runsGPU accelerated

Yes. Sarvam-105B runs on Apple M3 Ultra (256GB) at Q4_K_M (~67.5 GB of ~192 GB usable).

Needs ~67.5 GB Device usable ~192 GB

Runs at Q4_K_M using ~67.5 GB of ~192 GB usable. You have room for Q8_0 for higher quality.

Q4_K_M needed
~67.5 GB
Usable on device
~192 GB
Device memory
256 GB
Best quant
Q4_K_M

Run it

Install commands macOS

Pick your tool. All three load the same Q4_K_M weights.

llama.cpp
$ llama-cli -hf sarvamai/sarvam-105b-gguf:Q4_K_M
LM Studio
$ lms get sarvamai/sarvam-105b-gguf

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.

How to run it

On macOS use LM Studio (Polished GUI, ships MLX on Apple Silicon, one-click model downloads.). 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.

Model Sarvam
Parameters
105B (MoE, 10.3B active)
Q4_K_M size
64.2 GB
Context
128k
Full Sarvam-105B requirements →
Device macOS
Memory
256 GB unified
Usable for weights
~192 GB
Best runtime
MLX direct / Ollama (MLX backend)
Best models for Apple M3 Ultra (256GB) →

You could also run

Run Sarvam-105B on other hardware

FAQ

Can Apple M3 Ultra (256GB) run Sarvam-105B?

Yes. Sarvam-105B runs on Apple M3 Ultra (256GB) at Q4_K_M (~67.5 GB of ~192 GB usable).

How much memory does Sarvam-105B need?

Apple M3 Ultra (256GB) has room to spare. At Q4_K_M the weights are ~64.2 GB; with KV cache and runtime overhead, budget ~67.5 GB at a 4k context. It is a Mixture-of-Experts model (105B total / 10.3B active), so all experts must stay in memory; memory tracks total params, not active params.

What is the best tool to run Sarvam-105B 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.

Sources

Memory figures are estimates. See methodology.