model family

Sarvam

Published by sarvamai. 1 member has been measured on real machines. Every number below was signed by the machine that produced it — none of it is from a model card.

what it takes, what it returns
text → text

Declared by the network’s catalog, not measured — a signature says what this family would be asked. Whether it answered is the signed capability flags on each member.

1 of 4 builds measured · 3 still open
hf.co/bartowski/sarvam-1-GGUF:Q4_K_M1 · Q4_K_M · 3 machines · 2.53B
8,192recalled
—served
178.2best tok/s
23.5 tok/s · NVIDIA · Bhubaneswar13.0 tok/s · x86 · CPU · Hyderabad2.0 tok/s · NVIDIA · Bhubaneswar
2.0 → 23.5 tok/s across 3 machines · ×11.7 spreadNVIDIAx86 · CPU
recalled 8k⌐ declared 8k · unsigned

Builds this network knows of

Declared, not measured — this is what the catalog says exists and what it would cost a machine to hold. Nothing here is signed, and none of it claims the build works.

how much memory does your machine have

2 of 4 builds fit a 16 GB machine — budgeting 70% of it, or 11.2 GiB, because the cache and the runtime want the rest.

  • hf.co/bartowski/sarvam-1-GGUF:Q4_K_MQ4_K_Mollamameasured herequantized by bartowski
    1.4 GiB
  • sarvamai/sarvam-1BF16vllmpublished by sarvamai
    4.7 GiB
  • hf.co/bartowski/sarvamai_sarvam-m-GGUF:Q4_K_MQ4_K_Mollamaquantized by bartowski
    13.3 GiB
  • hf.co/sarvamai/sarvam-30b-gguf:Q4_K_MQ4_K_Mollamapublished by sarvamai
    18.2 GiB

Sizes are what the catalog declares, not what a machine measured — fitting on disk is necessary and not sufficient. Nobody here has run these on your machine, and this narrows the candidates rather than certifying one.

the same builds, by member
Sarvam-12B
hf.co/bartowski/sarvam-1-GGUF:Q4_K_MQ4_K_Mollama1.4 GiB on diskquantized by bartowski
sarvamai/sarvam-1BF16vllm4.7 GiB on diskpublished by sarvamai
Sarvam 30B (MoE)30B
hf.co/sarvamai/sarvam-30b-gguf:Q4_K_MQ4_K_Mollama18.2 GiB on diskpublished by sarvamai
Sarvam-M24B
hf.co/bartowski/sarvamai_sarvam-m-GGUF:Q4_K_MQ4_K_Mollama13.3 GiB on diskquantized by bartowski
ask your own gram

Your concierge can answer this from the network rather than from this page. The ask carries the operator id, so it resolves to a capability and an engine — not to a name somebody has to recognise.

What has this network measured about the Sarvam family (epn.family.sarvam), and can my gram run any of it? See /inferences/sarvam

open a gramx with this ask →epn.family.sarvam