mlx-community/context-1-MLX-4bit

Context-1 — MLX 4-bit

MLX quantization of chromadb/context-1 for Apple Silicon.

• Converted with mlx-lm version 0.31.2
• Also available: context-1-MLX-6bit

Key Specs

Detail	Value
Architecture	Mixture-of-Experts (MoE) Decoder-only Transformer
Base Model	gpt-oss-20b
Total Parameters	20B
Experts	32 routed, 4 active per token
Context Length	Up to 131,072 tokens
Attention	Alternating sliding window (128 tokens) + full attention
Quantization	4-bit affine, group size 64
Original Precision	BF16
Disk Size	~11 GB
Peak Memory	~12 GB
Chat Template	Supported

What is Context-1?

Context-1 is a 20B parameter agentic search model designed to retrieve supporting documents for complex, multi-hop queries. It works as a retrieval subagent alongside frontier reasoning models.

Key capabilities:

• Query decomposition — breaks complex multi-constraint questions into targeted subqueries
• Parallel tool calling — averages 2.56 tool calls per turn
• Self-editing context — prunes irrelevant documents mid-search (0.94 prune accuracy)
• Cross-domain generalization — trained on web, legal, and finance tasks

Performance: comparable to frontier LLMs at a fraction of the cost, up to 10x faster inference.

Performance on Apple Silicon

Metric	Value
Prompt Processing	227 tokens/sec
Generation	172 tokens/sec
Peak Memory	12 GB

Requirements

• Apple Silicon Mac with 16GB+ unified memory
• mlx-lm >= 0.31.2

pip install mlx-lm

Usage

CLI

mlx_lm.generate \
  --model mlx-community/context-1-MLX-4bit \
  --prompt "Your prompt here" \
  --max-tokens 256

Python

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/context-1-MLX-4bit")
response = generate(model, tokenizer, prompt="Your prompt here", max_tokens=256)
print(response)

LM Studio

This model is compatible with LM Studio on Apple Silicon. Search for context-1-MLX-4bit in the model browser and download directly.

Important: Agent Harness

Context-1 is designed to work with a specific agent harness that manages tool execution, token budgets, context pruning, and deduplication. The harness is not yet publicly released by Chroma. Running the model without it will not reproduce the reported benchmark results.

See the technical report for details on the agent harness design.

License

Apache 2.0

Credits

• Base model by Chroma
• MLX quantization by FF-01

Citation

@techreport{bashir2026context1,
  title = {Chroma Context-1: Training a Self-Editing Search Agent},
  author = {Bashir, Hammad and Hong, Kelly and Jiang, Patrick and Shi, Zhiyi},
  year = {2026},
  month = {March},
  institution = {Chroma},
  url = {https://trychroma.com/research/context-1},
}