saidutta69/FrogNano-4B-heretic

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 13, 2026License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

FrogNano-4B-heretic by saidutta69 is a 4.21 billion parameter decensored variant of Microsoft's FrogNano-4B-2609, a compact repository-level coding agent. Utilizing Heretic v1.4.0 for targeted weight edits, it suppresses refusal behavior while retaining the original's Gated DeltaNet architecture and Leaf five-tool harness. This model excels at software engineering tasks, achieving 61.5% on SWE-bench Verified, and is designed for local developer tooling, research agents, and CVE-triage workflows, particularly for English-language and Python-heavy repositories.

Loading preview...

Overview

FrogNano-4B-heretic is a 4.21 billion parameter coding agent developed by saidutta69, based on Microsoft's FrogNano-4B-2609. This model is a "decensored" variant, created using Heretic v1.4.0's directional ablation technique to suppress refusal behavior through targeted weight edits to the attention output and MLP down-projections, rather than traditional fine-tuning. While this interim checkpoint reduces refusals from 92/100 to 71/100, a stronger variant is planned. The model maintains its hybrid Gated DeltaNet architecture and Leaf five-tool harness, designed for repository-level software engineering tasks with an evaluated context of ~131K tokens.

Key Capabilities

  • Refusal Suppression: Significantly reduces refusal rates compared to the base model, making it more likely to attempt coding requests.
  • High Performance on SWE Benchmarks: Achieves 61.5% on SWE-bench Verified (up from 39.4% for its Qwen/Qwen3.5-4B starting point), 37.6% on SWE-bench Pro, 31.1% on Terminal-Bench 2.0, and 47.3% on PatchEval-Verified for CVE repair.
  • Local Execution: Optimized GGUF quantizations are provided, allowing it to run on various gaming PCs with 8GB to 24GB VRAM, or CPU-only.
  • Agentic Behavior: Designed as a purpose-built coding agent with a Leaf five-tool harness (read, write, edit files, glob-based search, shell execution).
  • Reproducible Abliteration: The full Heretic recipe is included for bit-for-bit regeneration of the model.

Good For

  • Local Developer Tooling: Ideal for developers seeking a coding agent that runs locally and provides direct answers without declining requests.
  • Research Agents: Suitable for research into autonomous software engineering agents.
  • CVE-Triage Workflows: Effective for identifying and repairing vulnerabilities within a sandbox environment.
  • English/Python Repositories: Validated for use with English-language and Python-heavy codebases.
  • Experimentation with Decensored Models: Useful for exploring models with suppressed refusal mechanisms, particularly in coding contexts.

Finetunes
1 models