---
license: apache-2.0
base_model: Qwen/Qwen3-Coder-30B-A3B-Instruct
language:
  - en
tags:
  - spring
  - spring-boot
  - java
  - backend
  - code
  - gguf
  - ollama
  - lora
  - peft
datasets:
  - bharathmuppa/spring-instruct
pipeline_tag: text-generation
library_name: transformers
---

# spring-coder-30b-a3b

A code model that is an expert in Spring Boot. It knows Spring Boot 3 and 4,
Spring Framework 6 and 7, Spring Security 6, Spring Data JPA, Hibernate,
MongoDB, Redis and Spring Cloud. It also knows Keycloak, Okta and Open Policy
Agent. It writes modern Java 21 and Spring code.
It was trained to write complete, compiling projects that follow one coding
guideline (`docs/SPRING-GUIDELINES.md` in the source repository).

- Base model: `Qwen/Qwen3-Coder-30B-A3B-Instruct`
- Method: LoRA (rank 16, attention projections q/k/v/o), merged into the base weights in bf16
- Trained with `transformers` and `peft` on one NVIDIA 80 GB GPU
- Data: `bharathmuppa/spring-instruct` (4613 instruction pairs: 4012 authored records on
  Spring Boot, Spring Security, Spring Data, Spring Cloud, Hibernate, Keycloak, Okta,
  Open Policy Agent, version facts and known issues, with 360 compiled project
  records, and 601 CodeAlpaca Java records in train only, CC-BY-4.0)

The weights are the full merged model in bf16, so this repository is large. The
adapter alone is in `bharathmuppa/spring-coder-30b-a3b-lora`.

## Use with transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
m = "bharathmuppa/spring-coder-30b-a3b"
tok = AutoTokenizer.from_pretrained(m)
model = AutoModelForCausalLM.from_pretrained(m, dtype="auto", device_map="auto")
msgs = [{"role": "system", "content": SYSTEM},
        {"role": "user", "content": "Write a SecurityFilterChain bean for Spring Security 6 that protects /api/** with JWT and leaves /actuator/health open."}]
text = tok.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False)
enc = tok(text, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(**enc, max_new_tokens=800)[0][enc["input_ids"].shape[1]:], skip_special_tokens=True))
```

The model was trained with this system prompt (`SYSTEM` above). Use it for best results:

> You are an expert Spring Boot engineer. You know Spring Boot 3 and 4, Spring Framework 6 and 7, Spring Security 6, Spring Data, Spring Cloud, Hibernate, Keycloak, Okta and Open Policy Agent. Answer with correct, modern, idiomatic Java 21 and Spring code and say which Spring Boot version a feature needs when it matters.

## Use with Ollama / llama.cpp (GGUF)

This repo also holds GGUF builds, made on the training pod with llama.cpp:
`spring-coder-30b-a3b-Q4_K_M.gguf` (about 18 GB, needs about 20 GB of memory) and
`spring-coder-30b-a3b-Q8_0.gguf` (about 32 GB), plus a `Modelfile` with the system prompt and
chat template.

```sh
ollama run hf.co/bharathmuppa/spring-coder-30b-a3b:Q4_K_M
```

Or build it locally from the Modelfile:

```sh
hf download bharathmuppa/spring-coder-30b-a3b --include "spring-coder-30b-a3b-Q4_K_M.gguf" "Modelfile" --local-dir spring-coder
cd spring-coder && ollama create spring-coder -f Modelfile && ollama run spring-coder
```

Q8_0 is closest to the bf16 weights; Q4_K_M is fine for scaffolding. On Apple
Silicon `mlx-lm` can also load the bf16 weights directly:

```sh
pip install mlx-lm
python -m mlx_lm convert --hf-path bharathmuppa/spring-coder-30b-a3b -q --q-bits 4 --mlx-path spring-coder-4bit
python -m mlx_lm generate --model spring-coder-4bit --prompt "Write a SecurityFilterChain bean that uses JWT."
```

## Training

| Setting | Value |
| --- | --- |
| Records (train / valid / test) | 4293 / 200 / 120 |
| Steps done | 1074 optimizer steps (2.0 epochs; the loop has a wall-clock limit of 180 minutes) |
| Final loss | train 1.056, eval 1.1449 |
| Batch | 1 x 8 gradient accumulation, max 4096 tokens |
| Learning rate | 3e-5, cosine, 50 warm-up steps |
| LoRA | rank 16, alpha 16, dropout 0.05, q/k/v/o |
| Loss | answer tokens only |

The run log with every loss line is in the dataset repository
`bharathmuppa/angular-lm-runs`. The dataset builder and the checker that compiles
the project records are in the `spring-lm` repository. The training pipeline
is in the GitLab project `mahaamatsya/ai/angular-lm`.

## Limits

The model writes scaffolds well. Logic can still have defects, so review the
output and run the compiler and the tests. Version facts after early 2026 come
only from the training data.

## Releases

Each Spring Boot version is a git tag on the Hub. The tag of this upload is
below. It marks the files of that version. Choose the tag in the branch menu, or add
`revision="<tag>"` when you download. The `main` branch holds the latest release.

This upload has no release tag.