AstraGPTCoder-7B

Abstract

A coding-specialised sibling of AstraGPT-7B. Same base, same think-tag reasoning format, tuned with LoRA for code generation. It runs in roughly 6GB of VRAM at 4-bit quantisation, which is what makes it usable on a single consumer GPU.

Overview

AstraGPTCoder-7B is the coding sibling of AstraGPT-7B. It shares the base model and the <think> reasoning format, and was fine-tuned with LoRA on code-heavy data to improve generation, completion and explanation of source code.

The practical difference is footprint: at 4-bit quantisation the model fits in roughly 6GB of VRAM, so it runs on a single 8GB consumer card without offloading.

Specifications

When to use it

• Code generation and completion where a local 7B is enough.
• Step-by-step explanation of unfamiliar code, using the think trace.
• Drafting tests and small refactors before a larger model reviews them.
• Not a replacement for execution: generated code still has to run and be reviewed.

Limitations

• Small-model code quality: complex repositories exceed its context and reasoning budget.
• The model card carries some older architecture notes from an earlier revision; treat AstraGPT-7B's card as the canonical configuration.
• English-primary, and no tool use out of the box.

Cite this

@misc{astragptcoder7b2026,
  author       = {Aditya Wakharkar},
  title        = {AstraGPTCoder-7B: A Coding-Specialised 7B LLM},
  year         = {2026},
  organization = {Tantra AI Labs},
  url          = {https://huggingface.co/tantra-ai-labs/AstraGPTCoder-7B}
}

The canonical version of this work lives on Hugging Face. If a number here and the card disagree, the card wins.