Summarization that follows your length.
Saransh (सारांश) is a compact, instruction-tuned language model built for faithful, readable summaries with controllable length. Summarize articles, reports, transcripts, research papers, and other English documents into one or two sentences, a concise paragraph, multiple paragraphs, or a specified word count.
With 1.72 billion parameters and quantized builds starting at approximately 1.1 GB, Saransh brings practical summarization to local machines without requiring a dedicated GPU.
Summaries need to fit their intended use.
A notification needs a sentence or two. A report preview needs a paragraph. A research brief needs enough detail to preserve the main findings. Saransh is fine-tuned to follow these length instructions more reliably than its untuned Qwen3-1.7B base model.
A Sentence or Two
For executive takeaways, headlines, push notifications, and document previews where immediate clarity and tight brevity are paramount.
A Concise Paragraph
For articles, email digests, meeting notes, reports, and research summaries that need balanced narrative context in a single cohesive block.
Preserving Main Findings
For detailed reports, research papers, and technical documents that require broader coverage without runaway verbosity or hallucinated claims.
Length Control That Works
Saransh is trained to adapt its output to your requested format rather than defaulting to an unnecessarily long response.
Length control that works
Choose a short, medium, or long summary, or specify an approximate word count. Saransh is trained to adapt its output to the requested format rather than defaulting to an unnecessarily long response.
Word budgets you can specify
Need a summary of approximately 120 words? Give Saransh an explicit word target. Around 35% of its curated training examples include numeric length instructions.
Summaries grounded in the source
Training data was filtered to remove examples containing unsupported numbers, named entities, and publication attributions. This helps reduce fabricated details and citations.
Compact enough to run locally
Quantized GGUF builds let you run Saransh on compatible laptops and local machines. Process documents locally without depending on a hosted inference API.
Open weights. Permissive license.
Saransh 1.7B is released under the Apache 2.0 license, inherited from its Qwen3-1.7B base model. Explore weights and integrate freely into applications.
One model. Four summary lengths.
Use the same source document and select the level of detail you need.
Global electrical transmission networks are undergoing an unprecedented modernization push as renewable energy sources exceed 30% of total generation across major industrial economies. Historically, regional power grids operated on predictable, centralized thermal generation schedules with dispatchable baseload reserves. Today, system operators face bidirectional power flows, volatile wind ramp rates, and decentralized battery assets distributed across thousands of commercial and utility nodes.
To stabilize grid frequency without relying on fossil-fuel peaking facilities, regional utilities are deploying grid-scale lithium-iron-phosphate (LFP) energy storage and synchronous condensers. Across North America and Western Europe, interconnection queues currently exceed 2,600 gigawatts of clean generation capacity, creating severe multi-year commercial bottlenecks. Regulators have responded by proposing dynamic line rating standards, updated transfer capacity metrics, and clustered study procedures to clear administrative backlogs.
Simultaneously, commercial data center expansion driven by high-throughput artificial intelligence workloads has introduced concentrated localized load spikes of 100 to 500 megawatts per facility campus. Hyperscalers are contracting directly with nuclear, hydro, and advanced geothermal operators for dedicated 24/7 clean baseload electricity, bypassing public utility interconnection timelines and accelerating the deployment of private corporate microgrid infrastructure. These converging pressures require grid operators and corporate energy buyers to balance real-time reliability against accelerated decarbonization targets.
Open weights. Permissive license.
Saransh 1.7B is released under the Apache 2.0 license, inherited from its Qwen3-1.7B base model. Explore the model weights, review the implementation, and integrate Saransh into compatible applications and workflows in accordance with the license.