Skip to content

Sarvam 1

Model Overview

Sarvam-1 is a 2-billion parameter language model specifically optimized for Indian languages. It provides best in-class performance in 10 Indic languages (bn, gu, hi, kn, ml, mr, or, pa, ta, te) when compared with popular models like Gemma-2-2B and Llama-3.2-3B. It is also competitive against the much larger models like Llama-3.1-8B in these languages. This is a text-completion model. It is meant to be finetuned on downstream tasks, and cannot be used directly as a chat or an instruction-following model.

Model Architecture

  • Hidden size: 2048
  • Intermediate size: 11,008
  • Number of attention heads: 16
  • Number of hidden layers: 28
  • Number of key-value heads: 8
  • Maximum position embeddings: 8,192
  • Activation function: SwiGLU
  • Positional embeddings: Rotary (RoPE) with theta=10,000
  • Training: Grouped-query attention and bfloat16 mixed-precision
  • Model Source: sarvamai/sarvam-1
  • License: Non-commercial License.

QPC Configurations

Precision SoCs / Tensor slicing NSP-Cores (per SoC) Full Batch Size Chunking Prompt Length Context Length (CL) Generated URL Download Generation Date
MXFP6 1 16 1 256 8192 https://dc00tk1pxen80.cloudfront.net/SDK1.20.4/sarvamai/sarvam-1/sarvam-1_qpc_16cores_256pl_8192cl_1fbs_1devices_mxfp6_mxint8.tar.gz Download 27-Feb-2026
MXFP6 4 16 1 256 8192 https://dc00tk1pxen80.cloudfront.net/SDK1.20.4/sarvamai/sarvam-1/sarvam-1_qpc_16cores_256pl_8192cl_1fbs_4devices_mxfp6_mxint8.tar.gz Download 27-Feb-2026

Run This Model

# Download QPC
mkdir -p sarvamai/sarvam-1
cd sarvamai/sarvam-1
wget <Download URL>
tar xzvf <downloaded filename.tar.gz>

# Run QPC
python3 -m QEfficient.cloud.execute --model_name sarvamai/sarvam-1 --qpc_path <path/to/qpc> --prompt "# shortest path algorithm\n" --generation_len 128

API Endpoint

# Start REST endpoint with vLLM
VLLM_QAIC_MAX_CPU_THREADS=8 VLLM_QAIC_QPC_PATH=/path/to/qpc python3 -m vllm.entrypoints.openai.api_server \
  --host 0.0.0.0 \
  --port 8000 \
  --model sarvamai/sarvam-1 \
  --max-model-len <Context Length> \
  --max-num-seq <Full Batch Size>  \
  --max-seq_len-to-capture <Chunking Prompt Length>  \
  --device qaic \
  --block-size 32