Saransh 1.7B
Summarization that follows your length.
A compact, instruction-tuned language model for faithful, controllable summarization.
Turn articles, reports, transcripts, papers, and documents into clear English summaries - from two sentences to detailed multi-paragraph outputs.
Small Model. Focused Purpose.
Saransh - सारांश, Sanskrit for “summary” - is a 1.7B parameter language model built specifically for summarization.
One to Three Pages
Paste one to three pages of text into Saransh. With an 8,192 trained token context window, articles, documents, and transcripts are analyzed as cohesive wholes without fragmentation.
Choose Your Detail
Choose how detailed you want the result to be, and Saransh produces readable English prose at the requested length — from tight two-sentence takeaways to comprehensive multi-paragraph briefs.
No GPU Required
With 4-bit quantization, the model occupies approximately 1.1 GB and runs comfortably on standard consumer laptops, MacBooks, and edge devices without needing a discrete GPU.
Core Advantages
Why Saransh?
Designed specifically to overcome the verbosity, hallucinated references, and resource overhead of untuned models.
Length Control That Actually Holds
Ask for one or two sentences and get one or two sentences.
In evaluation, the untuned base model produced an average of 252 words for short-summary requests. Saransh produced approximately 24 words.
Explicit Word Budgets
Need approximately 120 words? Just ask.
Explicit numeric targets are part of Saransh's trained behaviour rather than relying only on prompt compliance.
Built for Faithfulness
Training data was filtered to reduce unsupported information.
Examples were removed when reference summaries introduced numbers, named entities, or publication attributions absent from source documents.
Runs Locally
Quantized GGUF builds compatible with your workflow.
The Q4 build is approximately 1.1 GB, making local and offline summarization practical on everyday hardware without requiring cloud GPUs.
Open for Commercial Use
Released under the Apache 2.0 licence, inherited from Qwen3.
Use it for internal tools, research workflows, document processing, or commercial applications without licensing fees or usage restrictions.
One Model. Four Levels of Detail.
Choose how much information you want to keep.
The widespread adoption of foundation language models has transformed enterprise knowledge management, yet cloud-hosted models present persistent friction points: escalating inference costs, data residency vulnerabilities, and network round-trip latencies. To mitigate these obstacles, engineering teams are transitioning toward compact Small Language Models (SLMs) that execute directly on local CPU or edge hardware.
However, standard small models exhibit a notorious structural deficiency in summarization: an inability to adhere to length instructions. When prompted for a brief two-sentence executive brief, base models routinely produce rambling, multi-paragraph expositions averaging upwards of 250 words. In user-facing software, such unconstrained outputs shatter UI component boundaries, waste compute tokens, and dilute reader focus.
Saransh 1.7B addresses this limitation through targeted length-controlled alignment. Built on the Qwen3-1.7B architecture, the model underwent three successive data curation passes to eradicate phantom citations, normalize numeric entities, and couple explicit length buckets to distinct prompt phrasings. The resulting model delivers deterministic summaries matching the requested word count while maintaining strict factual alignment with the source text.
Built for Real Documents
Saransh was trained on a curated mixture covering the kinds of text people actually need to summarize.
News
Articles and multi-document news coverage.
Conversations
Everyday dialogue and messenger-style conversations.
Government & Legislation
Legislative documents and government research reports.
Scientific Research
Scientific and biomedical papers.
Long-Form Content
Literature and chapter-level documents.
The original 128,787 candidate examples were reduced to 97,038 curated training examples through multiple filtering passes.
Removed hallucinated publications, ungrounded numbers, and truncated texts before full-parameter fine-tuning.
Measured Improvement
Evaluated on 300 held-out documents spanning all ten source datasets.
| Metric | Qwen3-1.7B Base | Saransh 1.7B | Lift |
|---|---|---|---|
| ROUGE-1 | 0.190 | 0.387 | +104% |
| ROUGE-2 | 0.049 | 0.170 | +247% |
| ROUGE-L | 0.106 | 0.271 | +156% |
Tested with greedy decoding across held-out sets of all 10 datasets, reflecting massive gains in sentence coherence and semantic capture.
Model Specifications
Pick the Build That Fits Your Machine
Download quantized GGUFs directly or load in your preferred local LLM runtime.
Recommended default for laptops. Best balance between model size and output quality.
Slightly higher fidelity while remaining lightweight.
Near-lossless quantization for precision workloads.
Reference build with no quantization loss.
From Download to Summary in One Command
Run Saransh locally using Ollama:
ollama run hf.co/M37labsorg/Saransh-1.7B-GGUF:Q4_K_MOr load the model through LM Studio, llama.cpp, or the Hugging Face Transformers ecosystem.
# Run Saransh locally in one command:
ollama run hf.co/M37labsorg/Saransh-1.7B-GGUF:Q4_K_MDesigned for Document Workflows
Saransh is intended for summarizing English prose such as:
It works particularly well for:
Local and offline summarization, batch document processing, research tools, note-taking applications, document pipelines, and privacy-conscious deployments.
Know Its Limits
Saransh is a specialized summarization model — not a general-purpose assistant.
@misc{maithani2026saransh,
title = {Saransh: A Length Controlled Small Language Model for Abstractive Summarization},
author = {Maithani, Aniket for M37Labs},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/M37labsorg/Saransh-1.7B}}
}
Summarization Without the Infrastructure Overhead
1.7B parameters. ~1.1 GB quantized. No GPU required.
Bring controllable document summarization directly into your application, workflow, or local environment.