Chroma Context-1
by Chroma (no company note in the vault)
20B retrieval subagent: decomposes a hard query into several search steps, retrieves iteratively and prunes its own context to keep searching.
Facts (Chroma research page)
- Base: derived from gpt-oss-20B; 20B parameters, MXFP4 quantisation
- Release: 2026-03-26; weights under Apache 2.0, data-generation code also open
- Role: a search subagent used alongside a separate answering model, not a general chat model. The earlier wording “rewrites RAG queries” understated it: it runs multi-step retrieval and discards irrelevant documents as it goes.
- Claims (vendor): retrieval quality comparable to frontier LLMs at lower cost, up to 10x faster than larger models; results on generated web/finance/legal/email tasks and on BrowseComp-Plus and FRAMES.
- “Native ChromaDB integration” from the earlier note: unverified.
Pricing
Open weights, free to self-host.
Sources
- https://www.trychroma.com/research/context-1 (accessed 2026-09-30)
- https://huggingface.co/chroma/Context-1 (accessed 2026-09-30; card summary not obtained)