Chroma Context-1: Open-Weight 20B Search Agent Achieves Frontier-Level Retrieval at 10x Speed and 25x Lower Cost

Chroma / Hacker News·medium signal

Chroma released Context-1, a 20B-parameter agentic search model trained via SFT and RL from gpt-oss-20B. It decomposes queries into subqueries, searches iteratively, and self-edits its context window by pruning irrelevant documents (0.94 prune accuracy). Competitive with frontier models on BrowseComp-Plus, SealQA, FRAMES, and HLE benchmarks while being 10x faster and 25x cheaper. John Schulman endorsed the approach. Apache 2.0 licensed, weights on Hugging Face.