SecureCoder 30B Pro v3 — merged
LoRA fine-tune of unsloth/Qwen3-Coder-30B-A3B-Instruct
merged into full-weight safetensors, for code generation, tool calling and cybersecurity
(offence and defence).
Licence: Apache-2.0 — commercial use permitted, following the base model.
⚠️ Quality not yet verified
v3 was retrained after v1 and v2 were found to emit the literal two-character sequence
\n instead of real newlines, which made their Python fail ast.parse. The cause was
JSON-escaped message text in parts of the training mix; it is fixed in
Taimwe/securecoder-scripts.
Measured on the same prompts through the same harness:
| Model | valid Python |
|---|---|
base unsloth/Qwen3-Coder-30B-A3B-Instruct | 93.3% |
securecoder-30b-pro-v2 | 0.0% |
v3's own eval had not completed at the time of publishing. Treat the numbers above as
context, not as a claim about this checkpoint. See
HANDOFF.md.
Lineage
| Stage | Repo |
|---|---|
| Adapter (LoRA) | Taimwe/securecoder-30b-pro-v3 |
| Merged (this repo) | Taimwe/securecoder-30b-pro-v3-merged |
| Training / merge / quantise scripts | Taimwe/securecoder-scripts |
Architecture
| Architecture | Qwen3MoeForCausalLM (qwen3_moe) |
| Parameters | 30B total, ~3B active |
| Hidden size | 2048 |
| Context length | 262,144 |
| Precision | bfloat16, 13 shards (~56.9 GB) |
Quick start
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
MODEL = "Taimwe/securecoder-30b-pro-v3-merged"
tok = AutoTokenizer.from_pretrained(MODEL)
model = AutoModelForCausalLM.from_pretrained(
MODEL, torch_dtype=torch.bfloat16, device_map="auto")
messages = [{"role": "user", "content": "Write a race-free file watcher in Rust."}]
inputs = tok.apply_chat_template(
messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=512)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
A 30B MoE needs a capable GPU. For a smaller download use the Q4_K_M GGUF once published.
Limitations
- Produces working offensive-security code. Use only against systems you own or have written permission to test; you are responsible for its use.
- Inherits Qwen3-Coder's biases; no independent safety red-teaming was run.
- MoE inference stays memory-hungry even though only ~3B parameters are active per token.
- Behaviour outside the training mix is unverified.