ClaireChains

Reasoning Engine

Guide complex analysis with fewer tokens

Datarus-R1 mirrors human problem-solving by iterating through hypotheses, code execution, and decisive answers.

  • Learns from 144K ReAct-style analytical notebooks
  • Switch between Agentic and Reflection interfaces
  • Apache 2.0 licensing for unrestricted deployment
AtlasIDE Interface
Adaptive Reasoner

“Hand me your dataset or prompt, I will iterate and deliver a defensible answer.”

Token optimization
Token Savings 18-49%
Model Snapshot

Open weights for analytical excellence

Datarus-R1-14B-Preview is tuned from Qwen2.5-14B-Instruct to behave like a senior data analyst. It studies entire reasoning notebooks, capturing thought, action, and reflection loops to stay grounded while solving quantitative challenges.

Parameters
14.8B

Scaled to compete with 32B+ models while remaining cost-efficient.

Training Corpus
144K trajectories

Full analytical notebooks with reasoning, code execution, and self-corrections.

License
Apache 2.0

Commercial-friendly open weights ready for regulated environments.

Modes
Agentic + Reflection

Swap between ReAct planning and concise CoT write-ups on demand.

Quick Links
Explore the full Datarus ecosystem—from the hosted demo to production deployment assets.
WebsiteProduct roadmap, release notes, and contact.Try the DemoInteract with Datarus-R1 in the hosted chat experience.Hugging FaceDownload weights or serve via vLLM / TGI.Jupyter AgentNotebook pipeline with automatic retries and exports.Research PaperDive into the adaptive multi-step reasoning methodology.

Capabilities

From exploration to production

Datarus-R1 helps analysts, engineers, and operators make informed decisions with transparent reasoning steps.

Trajectory-Centric Training

Captures the "AHA" moments where hypotheses pivot, boosting accuracy on LiveCodeBench and AIME challenges.

Dual Interfaces

  • Agentic <step>/<action> planning
  • Reflection <think>/<answer> synthesis
  • Seamless mode switching mid-analysis

Enterprise Ready

Role-based controls, observability hooks, and private fine-tuning paths.

Performance

Benchmark results that rival bigger models

Trajectory-aware training keeps accuracy high while using up to 49% fewer tokens than peer models, enabling longer analyses within existing budgets.

BenchmarkDatarus-R1-14BEdge vs peers
LiveCodeBench v657.7+1.1 vs QwQ-32B
AIME 202470.1+17.5 vs DeepSeek-R1-Distill-14B
AIME 202566.2+3.1 vs Phi-4-reasoning
GPQA Diamond62.1+2.0 vs QwQ-32B
Dual Reasoning Modes

Choose between agentic and reflective flows

Toggle between interactive execution with tool use or compact reflections for documentation. Each mode preserves the ReAct-style tags adopted by Datarus notebooks.

Agentic (ReAct) Mode
Iterative planning with explicit <step>, <action>, and<observation> tags.
<step>
  <thought>Inspect recent credit risk failures</thought>
  <action>python_executor</action>
  <action_input>
data = load_latest_defaults()
plot_weekly_trends(data)
  </action_input>
</step>
Reflection (CoT) Mode
Summaries use <think> and <answer> tags for polished, auditable output.
<think>Adjusted ARIMA residuals confirm seasonality in charge-offs.</think>
<answer>The blended forecast improves MAE by 12% while keeping capital buffers stable.</answer>
Jupyter Agent Workflow
Built to operationalize Datarus-R1 with containerized notebook execution and exportable artifacts.

Multi-step execution

Orchestrate dockerized Jupyter notebooks, capture outputs, and feed observations back into the model.

Intelligent recovery

Automatic retries with error context keep analyses on track when imports fail or data drifts.

Rich artifacts

Generate shareable notebooks, structured transcripts, and business-ready narratives from a single run.

Integrations

Works where your teams operate

Embed Datarus-R1 inside notebooks, BI tools, or automation pipelines with SDKs and container images.

Marketplace Interface

Browse and deploy templates

Template 1
Rich Integrations library

Notebook Agents

Jupyter, VS Code, and Databricks ready.

Data Connectors

Snowflake, BigQuery, and S3 ingestion.

Deployment Targets

vLLM, TGI, TensorRT-LLM runtime support.

Modalities

Fluent across formats

Supports natural language, Python, SQL, and markdown with precise control instructions.

Natural Language

Supported

Python & Notebooks

Supported

SQL & Tabular

Supported

Markdown Reports

Supported

Workflow DSLs

Supported

Multilingual

Coming Soon
Expanded multilingual reasoning arriving this summer

Deploy anywhere

Serve on-prem or in the cloud with optimized runtimes for major GPU architectures.

vLLM

Runtime

TensorRT-LLM

Acceleration

TGI

Serving

DeepSpeed

Training

Ray

Distributed

Kubernetes

Orchestration

AWS / Azure / GCP

Cloud
Expanded MI325X optimized builds now in validation
Benefits

Why teams choose Datarus-R1

Reason reliably, audit effortlessly, and iterate quickly.

1

Transparent Reasoning

Structured traces make it easy to review every decision.

2

Operational Efficiency

Token savings reduce infrastructure cost while maintaining accuracy.

3

Ecosystem Ready

Open weights with Apache 2.0 licensing unlock commercial innovation.

Deploy Datarus-R1 where trust matters

Ready to deploy?

Partner with ClaireChains to productionize Datarus-R1 and ClaireAI side-by-side.

Datarus-R1 — Dual-mode reasoning LLM