eivintobias/heartly-qwen-code

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 7, 2026License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

eivintobias/heartly-qwen-code is a 1.5 billion parameter coding language model, fine-tuned from Qwen2.5-Coder-1.5B, featuring the Heartly hallucination-reduction architecture. It is specifically designed for multi-turn conversational code chat, incorporating a unique output grammar to separate reasoning from the final answer. This model aims to provide clean, accurate code responses by stripping internal grammar tags before presenting the output to the user, making it suitable for applications requiring reliable code generation.

Loading preview...

Heartly Qwen-Code v3: Hallucination-Reduced Coding LLM

This model, eivintobias/heartly-qwen-code, is a 1.5 billion parameter coding Large Language Model (LLM) built upon the Qwen2.5-Coder-1.5B base. Its core innovation lies in the Heartly hallucination-reduction architecture, which employs a unique output grammar to enhance reliability.

Key Capabilities & Features

  • Hallucination Reduction: Integrates a decide-verify-stop grammar, including thinking [reasoning] and <decide>speak|stop</decide><verify>known|unknown</verify> tags, to separate internal thought processes from the final user-facing answer. Only the clean [answer] is intended for the user.
  • Conversational Code Chat: Fine-tuned using a Stage-5 SFT recipe with 5,200 samples, specifically optimized for multi-turn conversational interactions related to code generation.
  • Grammar Stripping Utility: Provides reply_formatter.py and a FastAPI server.py to automatically process the model's raw output, stripping the internal Heartly grammar tags and presenting only the clean, formatted code answer.
  • Full Fine-tune: Represents a full fine-tune (not a LoRA adapter) of the base model, with the bottom 12 layers frozen during training.

Good For

  • Applications requiring reliable and less-hallucinatory code generation in a conversational context.
  • Developers looking for a small, efficient model (1.5B parameters) for code-related chat interfaces.
  • Use cases where clean, grammar-stripped code output is crucial, leveraging the provided utilities for post-processing.

Finetunes
1 models