Modelpublic

schoggie/Qwen3.6-35B-A3B-java-v1

sourceHugging Faceupdated 4mo agoView on Hugging Face

0likes29downloads

Model Card

Qwen3.6-35B-A3B-java-v1

QLoRA fine-tune of `Qwen/Qwen3.6-35B-A3B` targeting agentic Java coding and long-context positional recall.

GGUF quantizations live at `schoggie/Qwen3.6-35B-A3B-java-v1-GGUF`.

Training


Base	Qwen/Qwen3.6-35B-A3B (35B total / 3B active MoE, 256 experts, 8 active)
Method	QLoRA, r=32, target modules `q/k/v/o/gate/up/down_proj`
Trainable params	16.7 M (0.048 % of total)
Hardware	2 × A100 80 GB PCIe (RunPod), single-process model parallelism via `device_map="auto"`
Steps	528 (~10 h 17 min wall clock)
Final loss	1.39 → 0.27 (cosine schedule, mean token accuracy 87 %)
Dataset	4039 train + 51 eval rows, hybrid agentic-Java + positional-recall (DeepSeek + synthetic recall). Private.

Evaluation

40-example short-recall held-out set, compared against base Qwen3.6-35B-A3B at Q4KM on the same hardware:

Metric	Base Q4_K_M	Fine-tuned Q4_K_M	Δ
Avg grey/20	5.55	6.53	+18 %
Passed (≥ 8 grey)	11 / 40	14 / 40	+3 (+27 %)

Per-row breakdown: 13 improved, 15 tied, 12 worsened. The +3-pass go/no-go threshold for the pilot was met.

Intended use

—Java code generation in agentic loops (multi-step tool use, file edits, recall across large project context).
—Long-context Java retrieval — fine-tune emphasizes attending to relevant identifiers across ≥ 32 K-token windows.
—Drop-in replacement for the base in any Qwen3.6-A3B serving stack (llama.cpp, vLLM, sglang once supported).

Limitations

—Single-language fine-tune (English prompts, Java code). Other-language behaviour likely unchanged from base, untested.
—Eval was on a 40-row short-recall set; broader benchmarks (HumanEval-Java, SWE-bench, MMLU) were not re-run.
—Trained on Qwen 3.6 A3B specifically; not validated on the dense-7B variant or larger A22B sibling.

Inference

transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("schoggie/Qwen3.6-35B-A3B-java-v1")
model = AutoModelForCausalLM.from_pretrained(
    "schoggie/Qwen3.6-35B-A3B-java-v1",
    torch_dtype="auto",
    device_map="auto",
)

llama.cpp / Ollama / LM Studio

Use the GGUF repo: `schoggie/Qwen3.6-35B-A3B-java-v1-GGUF`. Q6_K is the production-deployed quant on the maintainer's dual-V100 box (27 GB, ~lossless).

Note on llama.cpp loader. Stock upstream llama.cpp has known loader bugs on the Qwen3.6-A3B GGUF metadata path. Use the unsloth-maintained fork until the upstream patch lands.

License

Inherits the Qwen Research License from the base model. The fine-tune adapter and merged weights are released under the same terms.

Citation

bibtex

@misc{qwen36-a3b-java-v1,
  author = {schoggie},
  title  = {qwen36-a3b-java-v1: Java-agentic QLoRA fine-tune of Qwen3.6-35B-A3B},
  year   = {2026},
  url    = {https://huggingface.co/schoggie/Qwen3.6-35B-A3B-java-v1}
}