vasishta/FastContext-1.0-4B-SFT-Fork

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 4, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

FastContext-1.0-4B-SFT-Fork by vasishta is a 4 billion parameter repository-exploration subagent, based on Qwen3-4B-Instruct, designed to optimize LLM coding agents. It specializes in issuing parallel read-only tool calls (READ, GLOB, GREP) to efficiently locate relevant code and return compact file paths and line ranges. This model significantly reduces main-agent token consumption and improves end-to-end resolution rates by providing focused context for coding tasks.

Loading preview...

FastContext-1.0-4B-SFT-Fork: An Efficient Repository Explorer for Coding Agents

This model is a 4 billion parameter supervised fine-tuned (SFT) variant of the FastContext-1.0 family, built upon the Qwen3-4B-Instruct backbone. It functions as a dedicated repository-exploration subagent for larger LLM coding agents, addressing the bottleneck of code exploration in complex development tasks.

Key Capabilities

  • Optimized Code Exploration: FastContext separates repository exploration from the main coding agent's task, allowing the main agent to receive clean, grounded evidence rather than raw search results.
  • Parallel Tool Calling: It can issue multiple READ, GLOB, and GREP tool calls in parallel to efficiently cover various search hypotheses within a repository.
  • Contextual Citation Generation: The model returns precise file paths and line ranges as focused context, significantly reducing the token budget consumed by the main agent.
  • Performance Improvement: Integration into coding agents like Mini-SWE-Agent has shown improvements in end-to-end resolution rates by up to 5.5% and reductions in main-agent token consumption by up to 60%.

Training & Tools

FastContext is trained using supervised fine-tuning on exploration traces, covering broad first-turn search, multi-turn evidence gathering, and precise citation generation. It exposes three read-only tools: READ for file contents, GLOB for path discovery, and GREP for regex searches.

Good For

  • Developers building or enhancing LLM-powered coding agents.
  • Reducing token usage and improving efficiency in code-related tasks requiring repository interaction.
  • Integrating a specialized subagent for intelligent code exploration and context provision.

Finetunes
1 models