Chroma drops Context-1 20B open-weights search agent

Context-1 signals a major shift toward tiered RAG architectures, proving that specialized, smaller models can outperform generalized frontier models at targeted tasks like search. The model actively self-edits its context window during multi-step searches to prevent "context rot," maintaining high precision over long horizons. It acts as a specialized retrieval engine that returns ranked documents to a downstream reasoning model, rather than generating the final answer itself. At 20B parameters and optimized for fast inference, it achieves GPT-4 level retrieval quality but is up to 10x faster and 25x cheaper. The release includes the model weights, agent harness, and full synthetic data pipeline under a permissive Apache 2.0 license.