---

base_model: georvn7/hayabusa-9b

library_name: transformers

pipeline_tag: text-generation

license: apache-2.0

language:

• en

tags:

• gguf
• llama-cpp
• text-generation
• debugging
• agent
• qwen3.5
• reinforcement-learning
• dpo

---

Hayabusa 9B GGUF

<p align="center">

<img src="assets/hayabusa9b.png" alt="Hayabusa 9B logo" width="260">

</p>

This repository provides GGUF quantizations for georvn7/hayabusa-9b.

Hayabusa 9B is a specialized, full-weight fine-tune of Qwen/Qwen3.5-9B created for autonomous software debugging and structured agent action selection. The model is trained on structured debugger contexts containing source code, stack traces, runtime logs, test outcomes, and execution history.

All GGUF files were converted from the original safetensors weights using llama.cpp at native BF16 precision, then quantized into standard k-quant variants.

---

Available Files and Quantizations

| File | Quant Type | Size | Description / Recommendation |

| :--- | :--- | :--- | :--- |

| hayabusa-9b-BF16.gguf | BF16 | 16.69 GB | Full precision base conversion. Highest fidelity reference weights. |

| hayabusa-9b-Q8_0.gguf | Q8_0 | 8.87 GB | Near-lossless 8-bit quantization. Recommended for critical debugging tasks. |

| hayabusa-9b-Q6_K.gguf | Q6_K | 6.85 GB | High quality retention with minimal degradation. Excellent balance. |

| hayabusa-9b-Q5_K_M.gguf | Q5_K_M | 6.02 GB | Good balance between memory usage and code analysis quality. |

| hayabusa-9b-Q4_K_M.gguf | Q4_K_M | 5.24 GB | Fast, lightweight 4-bit quant. Recommended default for local workstation inference. |

| hayabusa-9b-Q3_K_M.gguf | Q3_K_M | 4.31 GB | Compact footprint when GPU or system memory headroom is limited. |

| hayabusa-9b-Q2_K.gguf | Q2_K | 3.56 GB | Maximum compression for testing on low-memory hardware. |

---

Technical Specifications

• Base Model: georvn7/hayabusa-9b (fine-tuned from Qwen/Qwen3.5-9B)
• Checkpoint: hayabusa-9b-temporal-v2-step1-r1-c8-rl
• Architecture: Qwen3_5ForCausalLM (model_type: qwen3_5_text)
• Parameters: ~9B
• Attention Layers: 32 hidden layers (hybrid linear attention DeltaNet + full attention)
• Hidden Dimension: 4096
• Intermediate Dimension: 12288
• Context Length: 32,768 tokens training ceiling (up to 64K serving ceiling)
• Native Precision: bfloat16
• License: Apache 2.0

---

Training and Checkpoint Provenance

Hayabusa 9B represents the terminal checkpoint from Step 1, recovery Round 1, Cycle 8:

• Training Base: hayabusa-9b-temporal-v2-step1-r1-c7-rl
• Cycle Reference: Round 0 clean anchor
• DPO Rows: 47 (970,457 policy-branch tokens, train loss 0.6677)
• RL Rows: 42 (95 policy sequences, final approximate KL 0.000417)
• Reward Policy: Version 11
• Context Format: agent-optimized-debug-v3
• Assistant Thinking: Enabled and included in runtime generation contract

---

Prompt Format

Hayabusa 9B uses ChatML markup with thinking traces. The model generates internal reasoning steps inside <think>...</think> tags before emitting the concrete diagnosis or tool call:

<|im_start|>system
You are a software debugging agent specialized in diagnosing stack traces and selecting targeted fixes.<|im_end|>
<|im_start|>user
Function `find_needle(haystack, needle)` throws IndexError: list index out of range when needle is not present:

def find_needle(haystack, needle):

i = 0

while haystack[i] != needle:

i += 1

return i

Diagnose the root cause and provide the minimal fix.<|im_end|>
<|im_start|>assistant
<think>

---

How to Use

1. With llama.cpp

Run interactive generation with llama-cli:

llama-cli -m ./hayabusa-9b-Q4_K_M.gguf \
    -p "<|im_start|>system\nYou are a debugging agent.<|im_end|>\n<|im_start|>user\nDiagnose why this slice fails on empty inputs: return arr[0]<|im_end|>\n<|im_start|>assistant\n<think>\n" \
    -n 1024 \
    -c 4096 \
    --temp 0.6 \
    --top-p 0.95

Launch an OpenAI-compatible local server:

llama-server -m ./hayabusa-9b-Q4_K_M.gguf \
    --port 8080 \
    -c 8192 \
    --host 0.0.0.0

2. With Ollama

Create a file named Modelfile:

FROM ./hayabusa-9b-Q4_K_M.gguf

TEMPLATE """{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{- end }}
{{- range .Messages }}
<|im_start|>{{ .Role }}
{{ .Content }}<|im_end|>
{{- end }}
<|im_start|>assistant
<think>
"""

PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
PARAMETER temperature 0.6
PARAMETER top_p 0.95

Build and run the model in Ollama:

ollama create hayabusa-9b -f Modelfile
ollama run hayabusa-9b

3. With LM Studio

1. Copy the hayabusa-9b-GGUF directory or individual .gguf files to your LM Studio models directory (typically ~/.cache/lm-studio/models/).
2. Search for hayabusa-9b in the Local Models list.
3. Select ChatML as the prompt template.

---

Limitations

• Specialized for structured debugging and code action workflows rather than general conversational chat.
• Requires systematic prompt context (source code, logs, reproducible errors) for best results.
• Vision inputs are unsupported on this checkpoint (text-only fine-tune).

---

License

This model is licensed under Apache 2.0, inheriting upstream licensing from Qwen/Qwen3.5-9B. See the LICENSE file for details.