Context-1 is a 20B-parameter agentic search model designed as a specialized retrieval subagent to support multi-hop queries by handing results to a downstream frontier model.

The article provides open-source data-generation links and invites readers to explore further through the included URLs and social channels.

A core innovation is Self-Editing Context, which prunes irrelevant passages mid-search with a pruning accuracy of 0.94 to prevent context rot and keep retrieval quality within a 32k context window.

Chroma offers context-1-data-gen, an open-source tool that creates synthetic multi-hop tasks across four domains (Web, SEC filings, Patents, Email) using an Explore → Verify → Distract → Index pattern to emphasize reasoning over memorization.

Context-1 delivers up to 10x faster inference and roughly 25x lower cost than frontier models like GPT-5.4, while matching benchmark accuracy on HotpotQA and FRAMES via a four-agent parallel setup with reciprocal rank fusion.

The model builds on gpt-oss-20B, employing a Mixture of Experts and is fine-tuned with supervised learning and CISPO reinforcement learning to perform sequential reasoning and multiple parallel tool calls.

The release promotes a tiered RAG architecture where a fast subagent curates the golden context for a downstream frontier model, addressing latency and reasoning failures in large context windows.