Skip to content
AI Model Radar

FastContext 1.0 4B SFT

microsoft/FastContext-1.0-4B-SFT
Open Source · chat · open-weights · seen 12m ago
GA Alert me on changes
Context
—
Max output
—
Input $/1M
—
Output $/1M
—
Modalities
text
Released
14 Jun 2026
License: mit · microsoft/FastContext-1.0-4B-SFT
AI summary
● machine-written

Microsoft releases FastContext-1.0-4B-SFT, a 4B repository explorer for coding agents

FastContext-1.0-4B-SFT is a 4 billion parameter model designed to serve as a repository-exploration subagent for LLM coding agents. Based on Qwen3-4B-Instruct, it reduces token consumption in main coding agents by up to 60% while improving accuracy on benchmarks like SWE-bench Pro. The model performs parallel tool calls (READ, GLOB, GREP) to return focused code context without burdening the primary agent's token budget.

What's new
  • Up to 5.5% improvement on SWE-bench Pro when integrated with main agents
  • Up to 60.3% reduction in main-agent token consumption
  • Supports 262K token context length
  • Fine-tuned from Qwen3-4B-Instruct via supervised training
  • Designed for parallel read-only repository exploration
Best for
Building LLM-powered coding agents that navigate large code repositoriesReducing token consumption in multi-agent coding workflowsSoftware engineering tasks requiring precise code explorationScenarios requiring fast repository navigation without burdening primary LLM
Sources

Source: https://huggingface.co/microsoft/FastContext-1.0-4B-SFT