Instructions to use georvn7/hayabusa-9b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use georvn7/hayabusa-9b with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="georvn7/hayabusa-9b")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("georvn7/hayabusa-9b")
model = AutoModelForCausalLM.from_pretrained("georvn7/hayabusa-9b")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use georvn7/hayabusa-9b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "georvn7/hayabusa-9b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "georvn7/hayabusa-9b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/georvn7/hayabusa-9b

SGLang

How to use georvn7/hayabusa-9b with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "georvn7/hayabusa-9b" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "georvn7/hayabusa-9b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "georvn7/hayabusa-9b" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "georvn7/hayabusa-9b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use georvn7/hayabusa-9b with Docker Model Runner:
```
docker model run hf.co/georvn7/hayabusa-9b
```

Hayabusa 9B

Hayabusa 9B logo

Hayabusa 9B is a text-only, full-weight fine-tune of Qwen/Qwen3.5-9B for software debugging and agent-style action selection.

This is a full merged checkpoint, not a LoRA, QLoRA, or adapter release. The repository follows Hugging Face model conventions with root-level config.json, tokenizer files, model.safetensors.index.json, and sharded safetensors weights.

Important Behavior Notes

This model was trained as a text-only model.
Assistant thinking was disabled/removed in the no-assistant-thinking SFT views used for this checkpoint.
During inference, use the Qwen chat template with thinking disabled when supported.
The model is tuned toward debugging traces and action-oriented responses, not broad chat style.

Training Summary

The checkpoint is the result of continuation training stages, always continuing from the previous full checkpoint rather than restarting from the base model.

High-level lineage:

Full-weight SFT from Qwen/Qwen3.5-9B on the first no-assistant-thinking debugging dataset.
Round-2 continuation SFT on a larger no-assistant-thinking SFT union.
Round-2 DPO on cleaned debugging preference pairs.
Rare-actions SFT continuation.
Super-debug v3 main SFT continuation.
Super-debug v3 rare-actions SFT with final-assistant-message-only loss masking.
Super-debug v3 DPO continuation from the v3 rare-actions checkpoint.
Super-debug v3 analysis/system DPO continuation from the v3 DPO checkpoint.

Current uploaded checkpoint: 20260704_081724_hayabusa-9b_v3_analysis_dpo_393_from_hayabusa_16k_v1. Final train loss: 0.23098. Trained on 393 analysis-DPO preference rows. The final artifact is the merged bf16 full model, not an adapter.

Data

The first public dataset in this family is georvn7/super-debug-v1. Important caveat: that public dataset includes assistant thinking. This model was trained on no-assistant-thinking derived views for the SFT stages.

Current Capability Assessment

Recent Hen runs suggest Hayabusa-9B has learned a meaningful part of the debugger loop: it can follow structured actions, read runtime evidence, interpret progress reports, recover from some hangs, and make local fixes that move a failing project from missing artifacts to generated binaries. It is not yet a fully autonomous long-horizon debugger. In harder SimpleC compiler traces, the model can still get trapped in local repair loops when the correct fix requires multi-hop reasoning across parser AST construction, expression lowering, and generated assembly behavior. In those cases it may repeatedly patch the visible code-generation symptom instead of moving upstream to the true source of bad data. Hayabusa currently works best inside Hen with grounded traces, strict validation, progress reports, and preferably a stronger Director/reviewer model for long-context planning and fix review.

Loading With Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "georvn7/hayabusa-9b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": "What is the next debugging action for this trace?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
output_ids = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(output_ids[0], skip_special_tokens=True))

vLLM Serving Notes

Validated local serving profile on DGX Spark:

vllm serve georvn7/hayabusa-9b \
  --served-model-name hayabusa-9b \
  --max-model-len 65536 \
  --gpu-memory-utilization 0.70 \
  --max-num-seqs 1 \
  --max-num-batched-tokens 32768 \
  --dtype bfloat16 \
  --default-chat-template-kwargs '{"enable_thinking":false}' \
  --enforce-eager \
  --disable-frontend-multiprocessing \
  --language-model-only

Limitations

Specialized debugging/action model, not a general assistant benchmark release.
Not broadly safety-aligned beyond the underlying base model and task data.
Vision inputs are unsupported for this release.
Full reproducibility requires publishing the exact no-thinking SFT views and cleaned DPO files used in later stages.

License

This model inherits the upstream Qwen/Qwen3.5-9B Apache-2.0 license. See LICENSE.

Downloads last month: 652

Safetensors

Model size

9B params

Tensor type

BF16

Model tree for georvn7/hayabusa-9b

Base model

Qwen/Qwen3.5-9B-Base

Finetuned

Qwen/Qwen3.5-9B

Finetuned

(482)

this model