# Sarvam-105B: RAM and VRAM requirements

> Sarvam-105B is a 105B Sarvam model (Mixture-of-Experts, 10.3B active per token). At Q4_K_M it needs about **67.5 GB** to run and fits **3 of 39** tracked devices. Minimum to run: Apple M4 Max (128GB).

Last validated: 2026-06-15. Sources: Ollama, HuggingFace GGUF repos, vendor specs.

## Memory by quantization
| Quant | On disk | To run (4k context) |
| --- | --- | --- |
| Q4_K_M | 64.2 GB | ~67.5 GB |

Memory = weights + KV cache + ~0.8 GB runtime overhead, and varies ±15% with context length.

## Will it run on my device?
- **Apple M1 (8GB)** (8 GB): No, not enough memory
- **Generic Android Phone (8GB RAM)** (8 GB): No, not enough memory
- **iPhone 17 Pro** (12 GB): No, not enough memory
- **Nvidia GeForce RTX 4080 (16GB)** (16 GB): No, not enough memory
- **Google Pixel 10 Pro** (16 GB): No, not enough memory
- **Nvidia GeForce RTX 4090 (24GB)** (24 GB): No, not enough memory
- **Apple M4 Pro (48GB)** (48 GB): No, not enough memory
- **Apple M3 Ultra (256GB)** (256 GB): Yes, it runs — room for Q8_0

Full table of all 39 devices: https://localmodel.run/model/sarvam-105b

## How to run
Use LM Studio (Mac/Windows) or Ollama / vLLM (Linux).

## Details
- Parameters: 105B (MoE, 10.3B active per token)
- Default context: 128k tokens
- License: Apache-2.0 (commercial use: yes)
- Released: 2026-03
- HuggingFace: 18,743 downloads/mo, 275 likes

## FAQ
### How much VRAM or RAM does Sarvam-105B need?
About 67.5 GB at Q4_K_M (weights 64.2 GB + KV cache + overhead) at a 4k context.
### Can Sarvam-105B run on a laptop?
Sarvam-105B is large; you need a high-memory Mac or a 24 GB+ GPU at Q4_K_M.
### Can I use Sarvam-105B commercially?
Yes, Apache-2.0 permits commercial use.

Sources: https://huggingface.co/sarvamai/sarvam-105b, https://huggingface.co/sarvamai/sarvam-105b-gguf, https://www.sarvam.ai/blogs/sarvam-30b-105b
More: https://localmodel.run/model/sarvam-105b