Skip to content
localmodel.run

Text model · Sarvam

SSarvam-30B requirements

Sarvam family · 30B params (Mixture-of-Experts, 2.4B active) · released Mar 2026. Minimum to run at Q4_K_M: Nvidia GeForce RTX 4090 (24GB).

LicenseApache-2.0· Commercial OK↓ 38.8K/mo♥ 207on HuggingFace
Q4_K_M
19.6 GB
Q8_0
-
Total @ Q4 (4k)
~21.7 GB
Context
64 k

Quantization sizes

GGUF quantson disk
QuantizationSize on disk
Q2_K12.6 GB est
Q3_K_M14.7 GB est
Q4_K_M (default)19.6 GB
Q5_K_M21.4 GB est
Q6_K24.6 GB est
Q8_031.9 GB est
FP1660 GB est

Lower quant = smaller and faster, slightly lower quality. Q4_K_M is the common default.

Run it

llama.cpp
$ llama-cli -hf sarvamai/sarvam-30b-gguf:Q4_K_M
LM Studio
$ lms get sarvamai/sarvam-30b-gguf

Which devices can run Sarvam-30B?

FAQ

How much VRAM or RAM does Sarvam-30B need?

At Q4_K_M, Sarvam-30B needs about 21.7 GB (weights ~19.6 GB + KV cache + overhead) at a 4k context. At Q8_0 budget ~34 GB.

Can Sarvam-30B run on a laptop?

Sarvam-30B is large; you need a 24 GB+ GPU or a 32-48 GB Mac at Q4_K_M.

Is Sarvam-30B cheaper to run because it is a MoE model?

It is faster, not lighter. Sarvam-30B activates only 2.4B of 30B params per token (so it runs quickly), but all experts must stay in memory, so it still needs memory for the full 30B.

Can I use Sarvam-30B commercially?

Yes. Sarvam-30B is licensed Apache-2.0, which permits commercial use.

MoE with 128 sparse experts, top-6 routing, 2.4B active params. Released 2026-03 under Apache 2.0. Q4_K_M 19.6GB confirmed by summing the 6 shards in the official GGUF repo. No official Q8_0 GGUF released.

Sources

Memory figures are estimates. See methodology.