Jan-Nano-128k: Extended Context for Deep Research

Jan-Nano-128k, developed by Menlo Research, is a 4 billion parameter language model that significantly advances compact models for research. Building on Jan-Nano, this version features a native 128k context window, allowing for deeper and more comprehensive research capabilities without the performance degradation often associated with context extension techniques like YaRN.

Key Capabilities

Native 128k Context Window: Designed from the ground up to efficiently handle long contexts, maintaining performance across the full range.
Enhanced Performance: Unlike traditional methods, Jan-Nano-128k shows improved performance with longer contexts, as demonstrated on the SimpleQA benchmark.
Deep Research: Capable of processing entire research papers, lengthy documents, and complex multi-turn conversations.
MCP Compatibility: Maintains full compatibility with Model Context Protocol (MCP) servers.

Good For

Deep Document Analysis: Ideal for tasks requiring in-depth understanding of extensive texts.
Multi-Document Synthesis: Excels at combining and reasoning over information from multiple large sources.
Complex Reasoning: Suited for applications that demand intricate logical processing over large information sets.

This model is fully supported by Jan - beta build for seamless local AI experiences and can be deployed via VLLM with specific rope-scaling parameters to leverage its extended context.