amaara-code-explore
a code exploration agent
2 min readupdated
Amaara Code Explore is a specialized language model built on Qwen/Qwen3.5-4B-Base and trained to act as an agent for searching and analyzing repository source code. Given a bug report or question, the model explores the codebase using shell commands and file reads until it tracks down the relevant snippet. A key detail: it does not modify files or write patches, focusing strictly on navigation and pinpointing issues. This release is aimed at developers and engineers building autonomous coding systems.
Key features
- Built specifically for code navigation without file modification capabilities.
- Supports an extended context window of up to 262,144 tokens via quantization.
- Two-stage SFT (Supervised Fine-Tuning) focused on preventing repetitive loops.
- Ready-to-use GGUF quantizations from F16 to Q4_K_M for local deployment.
What you can do
- Hook the model up to a local server via llama-server to autonomously locate files from bug reports.
- Pair the model with mini-swe-agent for automated repository analysis.
- Test the model across long contexts to navigate large codebases.
Example
llama-server -m amaara-code-explore-Q8_0.gguf \
-ngl 99 -c 262144 --parallel 1 \
-ub 512 -b 1024 --jinja --host 0.0.0.0 \
--cache-type-k q8_0 --cache-type-v q8_0Read more
Details• Architecture: 4.66B parameters, base model Qwen3.5-4B-Base
• Metrics: 90% gold-file recall on a 10-task sample from SWE-bench Lite
• Limitations: does not generate patches or edit code, requires step-budget management
• License: Apache-2.0










Comments