Microsoft · 2026-06-15 · notable
FastContext — Microsoft's Explore subagent cuts coding-agent tokens by 60%
Microsoft and SJTU release FastContext: a small repository-exploration subagent that does parallel code search and hands focused context to a larger coding model. MIT-licensed code and 4B SFT and RL checkpoints are out.
Small 4B subagent that searches the repo for a bigger coding model, lifting SWE-bench resolution by up to 5.5% while cutting main-agent tokens by 60%.
Key specs
| GitHub stars | 361 |
|---|---|
| Resolution gain pct | 5.5 |
| Main token savings pct | 60 |
| Param range | 4B-30B |
Quick facts
| Makers | Microsoft Research and Shanghai Jiao Tong University |
|---|---|
| What it is | A small repository-exploration subagent that finds the right files for a larger coding model |
| License | MIT (code) |
| Explorer model sizes | 4B to 30B parameters |
| Training | Supervised fine-tuning followed by reinforcement learning |
| SWE-bench Verified gain | Up to 5.5% higher resolution |
| Token savings | 60% off the main model's tokens |
| Availability | Code on GitHub, 4B SFT and RL checkpoints on Hugging Face |
What is it?
FastContext is a research release from Microsoft Research and Shanghai Jiao Tong University. It splits coding-agent work into two roles: a cheap exploration model that finds the right files and snippets, and a stronger problem-solving model that does the actual reasoning over that focused context.
How does it work?
The explorer issues many code searches in parallel and returns short file-and-line evidence instead of dumping full files. The team trained explorers from 4B to 30B parameters with supervised fine-tuning followed by reinforcement learning, optimizing them to surface the right code with the fewest tokens. The main coding model then runs over only the explorer's distilled findings.
Why does it matter?
Coding agents hit a wall on long repos because context fills with irrelevant files. A specialised explorer raises SWE-bench Verified resolution by up to 5.5% while shaving 60% off the main model's token bill — useful for any agent that pays per-token. FastContext slots in as a callable tool for existing coding agents.
Who is it for?
coding-agent builders, SWE-bench researchers, anyone running multi-step LLM agents over large codebases
Frequently asked questions
- What is FastContext?
- FastContext is a research release from Microsoft Research and Shanghai Jiao Tong University. FastContext splits coding-agent work into two roles: a cheap exploration model that searches the repository and finds the right files and code snippets, and a stronger problem-solving model that reasons over that focused context. The smaller explorer hands distilled findings to the larger coding model.
- How much does FastContext improve coding agents?
- FastContext raises SWE-bench Verified resolution by up to 5.5% while cutting the main coding model's token use by about 60%. The explorer returns short file-and-line evidence instead of full files, so the main model reads only relevant context. These gains help any agent that pays per token when working over large codebases.
- Is FastContext open source?
- Yes. FastContext ships with MIT-licensed code on GitHub at github.com/microsoft/fastcontext. The team also released model checkpoints on Hugging Face, including 4B supervised fine-tuning and reinforcement-learning versions. FastContext is built to slot in as a callable tool for existing coding agents rather than as a standalone product.
- How does FastContext work?
- FastContext uses an explorer model that issues many code searches in parallel and returns short file-and-line evidence instead of dumping whole files. The team trained explorers from 4B to 30B parameters with supervised fine-tuning followed by reinforcement learning, rewarding them for surfacing the right code with the fewest tokens. The main coding model then runs over only those distilled findings.
Try it
https://github.com/microsoft/fastcontext