Question 1

How much VRAM does sarvam-105b need?

Accepted Answer

sarvam-105b (106B) requires approximately 278.43809509277344 GB of VRAM at full precision (FP16), 139.2 GB at 8-bit quantization (Q8), and 69.6 GB at 4-bit quantization (Q4). The actual VRAM usage may be slightly higher due to KV cache and runtime overhead.

Question 2

What GPUs can run sarvam-105b?

Accepted Answer

We have 1 compatible GPUs indexed for sarvam-105b, including MI350X. Compatibility is determined by whether the GPU's VRAM can fit the model's memory footprint, potentially using quantization to reduce memory usage.

Question 3

What is the cheapest GPU to run sarvam-105b?

Accepted Answer

The cheapest GPU to run sarvam-105b in the cloud is the MI350X at $5.49/hr.  Prices reflect the lowest on-demand rate across cloud providers.

Question 4

Can sarvam-105b run on consumer hardware?

Accepted Answer

sarvam-105b requires approximately 278.43809509277344 GB of VRAM at FP16, and even with 4-bit quantization (Q4) it needs about 69.6 GB. This exceeds the 24 GB available on the highest-end consumer GPU (RTX 4090), so sarvam-105b requires datacenter or cloud GPUs.

Question 5

What quantization options are available for sarvam-105b?

Accepted Answer

sarvam-105b supports the following quantization levels: FP16 (full precision): ~278.43809509277344 GB, Q8 (8-bit): ~139.2 GB, Q6 (6-bit): ~104.4 GB, Q4 (4-bit): ~69.6 GB. Lower quantization reduces VRAM usage at the cost of some quality loss. Q8 offers a good balance between quality and memory savings, while Q4 provides maximum compression. As a Mixture of Experts (MoE) model with 128 experts (8 active), sarvam-105b has a larger total weight size but only activates a subset of parameters per token, which can affect effective memory usage during inference.

sarvam-105b

VRAM Requirements

Compatible GPUs

sarvam-105b FAQ

VRAM requirements for sarvam-105b at different quantization levels
Quantization	VRAM Required
FP16	278.4 GB
Q8	139.2 GB
Q6	104.4 GB
Q4	69.6 GB