FrogNano-4B-2609

FrogNano-4B-2609

A compact coding agent model that works through a repository, finds bugs, edits files and runs tests with a five-tool framework called Leaf. Made for developers who want a small model for software issue fixing.

Made by microsoft on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: FrogNano is a compact coding agent model from Microsoft with 4.7B parameters. It is derived from Qwen3.5-4B and post-trained with reinforcement learning on about 1,500 synthetic software-engineering tasks. It works through a lightweight five-tool framework called Leaf, which runs its tool calls in an isolated copy of a repository.

Who it suits: It suits developers and researchers who want a small model that can read a codebase, locate a bug and propose a patch from a plain description.

Industry | Who uses it | How they use it | Example
Software | Developers fixing issues | Give the model a task description and let it work through the repository | Proposing a patch for a bug report
Research | Agent researchers | Study a small model trained with reinforcement learning only | Testing how it behaves inside the Leaf framework
Developer tools | Tool builders | Plug a compact coding model into an agent framework | Running tool calls against an isolated repository

Skip it if you need a model for images, video or non-code chat, since its extra training and evaluation are text only.

Where to use it:

• The Leaf framework, which the card names as the place its tool calls run
• Windows, macOS and Linux, with Python, to download the files
• A local machine with enough space for the weights

Use case | Who | What they get back
Resolve a software issue | Developers | A candidate patch for human review
Navigate and debug a codebase | Engineers | Cross-file understanding and bug fixes
Add a feature | Engineers | Multi-file edits refined with test feedback

How to set it up:

• Python, to run the download command
• About 9.3 GB of disk space for the BF16 weights, or about 2.3 GB at 4-bit as an estimate
• The Leaf framework, which runs the model's tool calls

1. Install the Hugging Face tools.
2. Download the model files to your computer.
3. Connect the model to the Leaf framework, following the model card and its linked repository.

pip install -U huggingface_hub
hf download microsoft/FrogNano-4B-2609

How to use it: You give the model a natural-language task description plus the existing source code. It replies with structured tool calls. Leaf carries them out in turn, searching files, changing code and running commands and tests, then returns the output to the model for the next round.

Tool | What it does
Leaf framework | Runs the model's tool calls in an isolated repository environment
Inspect and search files | Lets the model look through the repository
Edit code | Changes code to build a candidate patch
Run commands and tests | Runs them and returns the results to the model

Example: You describe a bug in a repository in plain English. The model inspects files, edits code and runs the tests through Leaf over several turns. What comes back is a candidate patch, which you review and test before using.

For newcomers and experts: Think of it as a small assistant for fixing code issues, one that needs its framework to act. A good first step is to download the model and read the card's section on the Leaf framework.

If you know the field, note that it inherits the base model's 32-layer design, which mixes Gated DeltaNet with gated attention. Training used executable test-based rewards over whole multi-turn coding runs. It never trained on stronger-model solutions. The evaluated setup uses about 131K tokens of context and up to 8,192 generated tokens per turn.

Task and code go to the model
Model writes structured tool calls
Leaf runs them in isolation
Test results return to the model
A candidate patch comes out

Good to know:

• Limits: performance depends on the Leaf framework and on test quality, and the training data are mostly Python and English.
• Review: patches can be wrong or insecure even when tests pass, so they need human review, regression tests and security checks.
• Modalities: image and video parts are inherited from the base but were not trained or evaluated.
• License: Apache License 2.0, with the base model's notices kept, although the page header lists MIT.
• Cost to run: none named.
• Data it reaches: it works on the repository you give it, through an isolated environment, and does not deploy changes itself.
• Not for: images or video, and any deployment without human review.

On Hugging Face, 13 November 2026

• Parameters: 4.7B
• Weights: about 9.3 GB in BF16; about 2.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 2,189
• Likes: 181
• License: apache-2.0
• Quantized versions: 24
• Fine-tunes: 2
• Tags: safetensors, qwen3_5

Links

• Model: https://huggingface.co/microsoft/FrogNano-4B-2609
• Files: https://huggingface.co/microsoft/FrogNano-4B-2609/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download microsoft/FrogNano-4B-2609

Then follow the usage in its model card. Nothing in this kit runs by itself.