Chroma Context-1

Chroma Context-1

A 20B-parameter search model that finds supporting documents for complex, multi-step questions and trims its own working context as it goes. Meant to act as a retrieval helper next to a larger reasoning model.

Made by chromadb on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Chroma Context-1 is a search model built on openai/gpt-oss-20b. It takes a hard question, splits it into smaller queries, searches a document collection step by step, and gathers the documents that support an answer.

It is meant to work as a retrieval subagent beside a larger reasoning model. While it searches, it drops documents it no longer needs, which leaves room in its context for more searching.

Who it suits: Developers who build retrieval or research agents and want a dedicated search step in front of a main reasoning model. It also interests researchers who read how agentic search is trained.

Industry | Who uses it | How they use it | Example
AI engineering | Agent builders | Add a search subagent that feeds a main model | A research agent that gathers sources before it answers
Legal | Legal tech developers | Retrieve supporting documents for multi-part questions | Finding documents that satisfy several stated conditions
Finance | Analytics teams | Search large document sets for related evidence | Collecting filings that bear on one question
Research | ML researchers | Study a self-editing search agent | Reading the technical report and comparing setups

Anyone who wants a model that runs from a few lines of code can wait, since it expects a specific agent framework that is not yet public.

Where to use it:

• A retrieval step inside an agent system, beside a larger reasoning model.
• Python with the Transformers library, since that is the library Hugging Face lists.
• Windows, macOS or Linux with Python, for downloading the files.

Use case | Who | What they get back
Multi-step retrieval | Agent developers | Supporting documents for a complex question
Query breakdown | Search engineers | A hard question split into targeted subqueries
Long searches | Research teams | Retrieval quality kept up by pruning irrelevant documents

How to set it up: You need Python and a machine that can hold a 20B model.

• Python with huggingface_hub
• Enough disk for the weights, about 41.8 GB in BF16 by Hugging Face's estimate
• A GPU or other hardware that can run a model of this size
• Chroma's agent framework, which is not yet public

1. Install the Hugging Face tool.
2. Download the model files.
3. Wait for or build a framework that manages tools, token budgets and pruning, as the card describes.

pip install -U huggingface_hub
hf download chromadb/context-1

How to use it: The card shows no run command. The model is trained for a specific framework that runs the tools, budgets tokens, prunes context and removes duplicates, and results reported in the technical report depend on it.

In day-to-day use the framework hands the model a query, the model decomposes it, calls search tools, often several at once, and returns the documents it found.

Example: A person gives the agent a question with several conditions that no single search can answer. Context-1 breaks it into separate subqueries, searches the corpus for each, drops documents that do not help, and returns supporting documents for the main model to use.

For newcomers and experts: If you are new, think of it as a research assistant that only finds sources. It does not write the final answer. The first small step is to read the technical report and download the files, because running it well needs the framework.

For experts, it is a mixture-of-experts model trained with supervised fine-tuning and then reinforcement learning with CISPO on a staged curriculum. The card reports an average of 2.56 tool calls per turn and a prune accuracy of 0.94. It says retrieval is comparable to frontier models at lower cost and up to 10 times faster, which are the authors' own claims.

A complex query arrives
The model splits it into subqueries
It searches the corpus with parallel tool calls
It prunes documents that do not help
Supporting documents go to the main model

Good to know: The points below cover limits and costs.

• Framework: running the model without the agent framework will not reproduce the reported results, and that framework is not yet public.
• Precision: the files are BF16, and a smaller MXFP4 checkpoint is listed as coming soon.
• License: Apache 2.0.
• Cost to run: no API key or paid service is named. A 20B model needs substantial hardware.
• Data it reaches: it searches the document corpus your framework gives it, and nothing else is named.
• Not for: writing final answers on its own, or standalone use without the framework if you want the reported quality.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 20.9B
• Weights: about 41.8 GB in BF16; about 10.5 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 613
• Likes: 425
• License: apache-2.0
• Based on: openai/gpt-oss-20b
• Quantized versions: 8
• Tags: transformers, safetensors, gpt_oss, text-generation, conversational

Links

• Model: https://huggingface.co/chromadb/context-1
• Files: https://huggingface.co/chromadb/context-1/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download chromadb/context-1

Then follow the usage in its model card. Nothing in this kit runs by itself.