Chroma Context-1: Open-Weight 20B Search Agent That Edits Its Own Context

Chroma, the open-source vector database company, has released Context-1 — a 20B parameter agentic search model trained with reinforcement learning that achieves retrieval performance comparable to frontier LLMs at 10x faster speed and 25x lower cost.

Context-1 is designed as a subagent for search tasks. Given a query, it decomposes it into subqueries, iteratively searches a corpus, and returns a ranked set of supporting documents. The key innovation is self-editing context: the agent actively decides which retrieved information to retain and which to discard, freeing capacity for further exploration within a bounded context window.

The model was trained on over 8,000 synthetically generated agentic search tasks using RL, starting from gpt-oss-20b. It operates at the Pareto frontier of cost, latency, and quality — making it practical for production deployments where frontier model search would be prohibitively expensive.

Chroma has released the full open weights on HuggingFace (https://huggingface.co/chromadb/context-1) along with the complete data generation pipeline for reproducibility. The research blog is at https://trychroma.com/research/context-1.

For the agentic ecosystem, Context-1 represents a new category: purpose-built, RL-trained subagents optimized for specific agent capabilities like search, rather than general-purpose LLMs repurposed for tool use.