Chroma Context-1 — 20B-parameter open-source agentic search model (Apache 2.0)

TL;DR

Chroma announces Context-1, a 20B-parameter “search agent” model released under Apache 2.0. The claims in the launch tweet are quantitative on three axes: a new Pareto-frontier position for agentic search; ~10× faster than incumbents; ~10× cheaper. No paper, benchmark numbers, or technical report are included in the tweet itself — it is a product/release announcement with an attached demo video.

Key claims

Context-1 is a 20B-parameter “search agent” model [tweet].
Chroma claims it “pushes the Pareto frontier of agentic search” — i.e. a new dominance point on cost × latency × quality vs. existing agentic search systems [tweet].
Order-of-magnitude faster than prior systems (1×→10× class) [tweet].
Order-of-magnitude cheaper than prior systems [tweet].
Released under Apache 2.0 — fully open-source, weights presumably included [tweet].

Method

Not disclosed in the tweet. Chroma is best known as a vector database; “agentic search” in their framing typically means a model that drives retrieval (query reformulation, multi-hop, tool-augmented) rather than acting as a passive embedding model. 20B parameters places it between recent open agentic-search baselines (32B class in InfTool / ReTool / MiroThinker-32B) and below the 70B+ frontier-comparable open models. No training-recipe details, base-model lineage, or benchmark suite are stated in the tweet.

Results

No quantitative benchmark numbers in the announcement — the claims are stated in unitless “order of magnitude” form. Comparison points (which incumbents, on which benchmark, at which agent-loop budget) are not specified. The attached media is a video demo (image asset only).

Why it’s interesting

Context-1 lands in a specific gap on the Tool-Use Agents design space: the parameter-space camp (InfTool, MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling) has been operating at 32B–72B with full GRPO post-training, while the context-space camp (Training-Free Group Relative Policy Optimization) keeps a frozen frontier model and learns a natural-language experience library. A 20B Apache-2.0 model with explicit “order-of-magnitude faster + cheaper” claims would, if validated, be a new datapoint on the cost-vs-capability frontier of Training-Free Group Relative Policy Optimization‘s Figure 1 — but with parameter-space tuning rather than context-space. It is also another open multi-purpose release in the cohort tracked by Open foundation-model releases from a vector-DB vendor (a vertically-integrated retrieval+model release pattern that is rare so far on this wiki — the closest analog is Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking‘s embedding+reranker pairing, but Context-1 appears to be a single agent model rather than a pipeline pairing).