DJLougen / Qwable-5-27B-Coder

Qwable-5-27B-Coder

Qwable-5-27B-Coder is a Qwen3.6-based coder-agent tune trained first on Claude Fable 5 traces , then continued on Kimi 2.7 Coder traces . It is built for the messy part of coding work: reading a repo, planning a patch, using terminal feedback, fixing the miss, and carrying constraints through long turns.

Early maintainer runs show Qwable outperforming the base model on a private coder benchmark. Public scores, harness settings, and task definitions will be added when the evaluation packet is ready.

Training, quantization, and coder-agent evaluation are expensive. If Qwable helps your work, support continued releases at ko-fi.com/djlougen .

Release channels

Trace stack

unsloth/Qwen3.6-27B
  -> Claude Fable 5 coder-agent traces
  -> Kimi 2.7 Coder traces
  -> Qwable-5-27B-Coder

The release is aimed at agentic coding behavior, not benchmark-demo prose. The training signal is trace-shaped: inspect, decide, edit, verify, recover.

What Qwable is tuned to do

• Navigate real repositories instead of isolated snippets.
• Translate failing command output into the next useful patch.
• Keep constraints alive across multi-step coding tasks.
• Produce tool-friendly, implementation-oriented answers.
• Handle long engineering prompts with logs, diffs, stack traces, and partial failures.
• Bias toward concrete edits, commands, and verification over generic advice.

Quickstart

Install a recent Transformers build that supports the Qwen3.6 / Qwen3-VL model family. The checkpoint is large; use device_map="auto" or an equivalent sharded serving setup.

pip install -U transformers accelerate safetensors pillow

Text-only coding prompt:

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "DJLougen/Qwable-5-27B-Coder"

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are Qwable, a precise coding agent. Inspect first, patch carefully, verify behavior."},
    {"role": "user", "content": "Write a Python function that merges overlapping intervals, then explain the edge cases."},
]

prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[prompt], return_tensors="pt").to(model.device)

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        max_new_tokens=1024,
        do_sample=True,
        temperature=1.0,
        top_p=0.95,
        top_k=20,
    )

new_tokens = output_ids[:, inputs.input_ids.shape[-1]:]
print(processor.batch_decode(new_tokens, skip_special_tokens=True)[0])

Download locally:

hf download DJLougen/Qwable-5-27B-Coder --local-dir Qwable-5-27B-Coder

Prompting profile

Qwable works best when the prompt looks like an actual coding task, not a riddle.

Good inputs include the relevant files, exact failing command output, hard constraints, expected output format, tool boundaries, and a verifier command or acceptance test when available.

Suggested system prompt:

You are Qwable, a precise coding agent. Inspect before editing. Prefer minimal, correct patches. Preserve existing conventions. Verify behavior with the narrowest meaningful test before finalizing.

For benchmark runs, keep prompts, sampling, max tokens, and tool schema exposure identical between the base model and Qwable. The current generation_config.json uses temperature=1.0 , top_p=0.95 , and top_k=20 .

Evaluation status

Current public status: early maintainer testing only. The maintainer has observed wins over the base model on a private coder benchmark, but reproducible claims require the full packet: benchmark name, split, prompt format, tool schema, harness commit, sampling settings, pass/fail rules, and raw results.

Vision and multimodal note

The repository is configured as image-text-to-text , and the base model family supports image/video tokens through the Qwen vision stack. This fine-tune is marketed for coding behavior. Do not assume it improves vision understanding unless you evaluate that separately.

Limitations

• Public benchmark scores are not published yet.
• The model may inherit failure modes from the base model and from the trace sources.
• Long-context behavior depends on runtime implementation, hardware, KV cache settings, and prompt structure.
• Tool-use quality depends on prompt format and schema consistency.
• This is a large BF16 checkpoint; most local users will need quantization or multi-GPU serving.
• The card does not claim safety alignment beyond the base model and fine-tuning data.

License

Released under Apache-2.0, following the upstream base model license metadata.