Amaara Code 4B v8
Specialized AI Model for Coding and Development
2 min readupdated
Amaara Code 4B v8 is a model designed for programming and agentic workflows, fine-tuned on a unified dialogue corpus shared across the entire Amaara Code line. This repository provides quantized weights in the GGUF Q8_0 format (8-bit quantization, file size 4.61 GB) for running via llama.cpp. The model uses the standard Qwen ChatML prompt template and is built for developers who need a compact tool for writing, debugging, and reviewing code locally or on resource-constrained servers.
Key features
- Base architecture: Built on top of Qwen/Qwen3.5-4B-Base, retaining the shared training dataset of the Amaara Code line.
- Quality data: The corpus is filtered by trajectory shape (retaining only verified "edit — test — pass" chains and self-correction steps).
- GGUF format: The Q8_0 version weighs 4.61 GB and is optimized for efficient inference in llama.cpp.
- Licensing: Distributed under the A-TECH Model Licence in addition to the upstream Apache-2.0.
What you can do
- IDE integration: Connect the model to local development environments or plugins via llama.cpp for autocompletion and function generation.
- Agentic tasks: Use the model as an engine for local coding assistants capable of step-by-step debugging.
- Code analysis: Apply it to bug hunting and refactoring small blocks of code across various programming languages.
Example
The model runs through compatible clients and llama.cpp-based backends. Use the Qwen ChatML prompt template to send requests for code generation or debugging.
Read more
Details• Model weight: 4.61 GB (Q8_0 GGUF format)
• Architecture: 4B parameters, based on Qwen/Qwen3.5-4B-Base
• Evaluation: tested using A-Coder-CLI on an SWE-bench-inspired task suite










Comments