microsoft/FrogNano-4B-2609

VISIONPricing:Input $0.1 / Cached $0.08 / Output $0.2Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 17, 2026License:mitArchitecture:Transformer0.1K Open Weights Featherless Exclusive Cold

FrogNano-4B-2609 by Microsoft is a 4.5 billion parameter repository-level coding agent, derived from Qwen/Qwen3.5-4B. It is post-trained exclusively with reinforcement learning on approximately 1,500 synthetic software-engineering tasks, utilizing a 32-layer hybrid Gated DeltaNet/gated-attention architecture and a 131K token context length. This model excels at long-horizon repository navigation, debugging, code editing, and patch generation, operating through the five-tool Leaf harness for iterative software issue resolution.

Loading preview...

FrogNano-4B-2609: A Specialized Coding Agent

FrogNano-4B-2609, developed by Microsoft, is a 4.5 billion parameter language model specifically designed as a repository-level coding agent. Derived from Qwen/Qwen3.5-4B, it features a dense 32-layer hybrid Gated DeltaNet/gated-attention architecture. Unlike its base model, FrogNano is post-trained exclusively with reinforcement learning on approximately 1,500 synthetic software-engineering tasks, generated and calibrated using TaskPilot. This specialization allows it to perform long-horizon repository navigation, debugging, code editing, and patch generation without relying on stronger-model solution trajectories.

Key Capabilities

  • Repository-level issue resolution: Addresses software issues from natural-language descriptions and existing source code.
  • Codebase understanding and manipulation: Facilitates navigation, cross-file code understanding, debugging, bug fixing, and feature implementation.
  • Structured tool calls: Generates Leaf tool calls for inspecting files, editing code, running shell commands, and executing tests within an isolated environment.
  • Iterative patch generation: Supports multi-file patch generation and refinement using command and test feedback.
  • Long-context processing: Handles combined interaction contexts of approximately 131K tokens.

Good For

  • Software-engineering research: Ideal for studying repository-level coding agents.
  • Human-supervised development: Assists developers with bug diagnosis, repair, scoped feature implementation, and test-driven code maintenance.
  • Python-heavy repositories: Best suited for English-language, Python-centric codebases with reproducible environments and executable test suites.
  • Controlled environments: Intended for use within sandboxed environments like the Leaf harness, with human review and testing of generated patches.

Finetunes
1 models