DroidSenior-7B-v5
a model built for modern Android and Kotlin development
2 min readupdated
A specialized LLM for writing modern Android code in Kotlin (Jetpack Compose, Coroutines/Flow, Hilt, Room). Created by celebibasar on top of Qwen/Qwen2.5-Coder-7B-Instruct (7.62B parameters), it was trained in a single run from a clean checkpoint rather than continuing from earlier versions. The model targets stable library versions as of November 2026, fixes errors using compiler output, and avoids deprecated APIs. It is a solid fit for Android developers generating clean code or migrating existing projects.
Key features
- Up-to-date stack (November 2026): trained and tested on AGP 9.5.1, Kotlin 2.4.21, Gradle 9.9.0, compileSdk 37, and Compose BOM 2026.09.00.
- Compiler-feedback reinforcement learning (GRPO): trained for 150 RL steps with rewards tied to clean builds free of deprecation warnings.
- Multi-turn error fixing: accepts compiler error logs in follow-up messages to patch its own code and respects project import hints.
- Lightweight local deployment: available as a 4.68 GB Q4_K_M GGUF quantization that runs even on a CPU.
What you can do
- Generate Jetpack Compose screens and Room/Coroutines data layers without deprecated methods or over-engineered architecture.
- Plug it into an automated repair loop by feeding Gradle or compiler output back into the model prefixed with
Checking your answer against the project found:to fix bugs automatically. - Pass your project's current imports (
Relevant APIs in this project) in the prompt so the model uses the exact library paths right away.
Example
System prompt from the model card to set the assistant's role:
You are DroidSenior, a staff-level Android engineer. Write idiomatic, modern Kotlin (Jetpack Compose, Coroutines/Flow, Hilt, Room) following the official Android architecture guidance. Explain key design decisions and trade-offs briefly, never use deprecated APIs, avoid over-engineering, and include tests when they add value. Answer in the user's language.
Read more
DetailsBenchmark (368 held-out tasks): pass@1 jumped from 46.2% (v4) to 67.9% (post-GRPO), reaching 82.3% with two rounds of compiler-log fixes.
Training: SFT on 7,614 examples (LoRA r=64, alpha=128) distilled from Qwen2.5-Coder-32B-Instruct, followed by 150 steps of GRPO.
Resources: the Q4_K_M GGUF build is 4.68 GB and runs at 7.9 tokens/s on a Google Colab CPU.
Limitations: library knowledge cuts off on November 8, 2026, and generated code should still be reviewed before use.










Comments