aravindpersona/restraint-7b

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 12, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Restraint-7B by aravindpersona is a 7.6 billion parameter tool-use model, based on Qwen/Qwen2.5-7B-Instruct, specifically post-trained to enhance tool-calling precision. It excels at answering knowledge questions directly without unnecessary tool calls, while accurately invoking tools when required. This model addresses the common issue of tool-capable LLMs over-tooling, significantly reducing unnecessary tool calls and improving overall accuracy on complex tasks.

Loading preview...

Restraint-7B: Precision Tool-Use for Knowledge and Action

Restraint-7B is a 7.6 billion parameter language model, built upon the Qwen/Qwen2.5-7B-Instruct base, that has been meticulously post-trained to optimize its tool-use behavior. Unlike many tool-capable models that tend to over-tool, Restraint-7B is designed to answer knowledge-based questions directly when it possesses the information, and to precisely call tools only when they are genuinely needed for complex tasks.

Key Capabilities

  • Intelligent Tool Restraint: Achieves a 94.1% restraint rate on no-tool tasks, meaning it avoids unnecessary tool calls for questions it can answer directly.
  • Enhanced Accuracy: Demonstrates a significant improvement in task accuracy, rising from 78% (baseline) to 92% on a 50-task stratified evaluation.
  • Reduced Tool Calls: Decreases total tool calls by 57% compared to its baseline, leading to lower latency and token usage.
  • Robust Multi-step Reasoning: Successfully handles Tier 1 single-tool and Tier 3 multi-step tasks with high reliability.
  • Specialized Training: Fine-tuned using 450 expert ReAct traces and a corrective SFT pass to refine its decision-making process for tool invocation.

Good For

  • Applications requiring efficient and precise tool-use, where minimizing unnecessary API calls is crucial.
  • Building agents that need to differentiate between direct knowledge retrieval and tool-assisted problem-solving.
  • Scenarios where a balance between direct answers and complex tool-chaining is desired, such as in customer support bots or intelligent assistants.
  • Developers looking for a 7B model that offers improved performance and resource efficiency in tool-augmented LLM workflows.

Finetunes
1 models