Instructions to use anandpranav/sarvam-20pct-distilled-resilient2026 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use anandpranav/sarvam-20pct-distilled-resilient2026 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="anandpranav/sarvam-20pct-distilled-resilient2026", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("anandpranav/sarvam-20pct-distilled-resilient2026", trust_remote_code=True, dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use anandpranav/sarvam-20pct-distilled-resilient2026 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "anandpranav/sarvam-20pct-distilled-resilient2026"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "anandpranav/sarvam-20pct-distilled-resilient2026",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/anandpranav/sarvam-20pct-distilled-resilient2026

SGLang

How to use anandpranav/sarvam-20pct-distilled-resilient2026 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "anandpranav/sarvam-20pct-distilled-resilient2026" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "anandpranav/sarvam-20pct-distilled-resilient2026",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "anandpranav/sarvam-20pct-distilled-resilient2026" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "anandpranav/sarvam-20pct-distilled-resilient2026",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use anandpranav/sarvam-20pct-distilled-resilient2026 with Docker Model Runner:
```
docker model run hf.co/anandpranav/sarvam-20pct-distilled-resilient2026
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Sarvam-30B — 20% Pruned + Knowledge Distilled (26.4B Parameters)

Overview

This repository contains a deployment-ready 20% pruned and knowledge-distilled version of Sarvam-30B (Mixture-of-Experts) prepared for the Resilient AI Challenge 2026 (Text-to-Text Track).

The model was compressed using structured expert pruning followed by Knowledge Distillation (KD) to recover accuracy lost during pruning.

Performance Summary

Metric	Baseline (32.15B)	This Model (26.4B)	Difference
Accuracy	95.12%	85.37%	-9.75%
Relative Accuracy	100%	89.75%	-10.25%
Throughput	7.72 tok/s	7.90 tok/s	+2.3% faster
Latency	129.78 ms/tok	126.74 ms/tok	-2.3% lower
Energy (41 samples)	83.25 Wh	81.30 Wh	-2.4% less
Disk Size	61.44 GB	49.14 GB	-20.0% smaller
Parameters	32.15B	26.4B	-18.0% fewer

Evaluation Dataset: 41 Indic Language Samples (Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati)

Compression Method

Step 1 — Structured Expert Pruning (20%)

Removed 20% of MoE expert layers (578 experts removed)
Pruning based on expert activation frequency
Result: 32.15B to 26.4B parameters

Step 2 — Knowledge Distillation

Teacher: Original Sarvam-30B (32.15B)
Student: 20% pruned Sarvam (26.4B)
8 epochs of Knowledge Distillation
280M LoRA parameters (1.13% of model)
Training loss reduced by 53%
Accuracy recovered: +2.44% (from 82.93% pruned-only to 85.37%)

Inference

Launch with vLLM

vllm serve --config vllm_config.yaml

Downloads last month: 203

Safetensors

Model size

26B params

Tensor type

BF16

Model tree for anandpranav/sarvam-20pct-distilled-resilient2026

Base model

sarvamai/sarvam-30b

Finetuned

(10)

this model