---
license: apache-2.0
base_model:
  - Qwen/Qwen3.5-9B
  - unsloth/Qwen3.5-9B
  - darkc0de/Qwen3.5-9B-heretic
  - lukey03/Qwen3.5-9B-abliterated
  - llmfan46/Qwen3.5-9B-ultimate-irrefusable-heretic
  - llmfan46/Qwen3.5-9B-ultra-heretic
  - jwest33/qwen3.5-9b-null-space-abliterated
  - osirisbrain/OsirisCortex-v6
  - DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-THINKING
  - DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT
  - crownelius/Crow-9B-Opus-4.6-Distill-Heretic_Qwen3.5
  - alecccdd/Qwen3.5-9B-paraphrasing-orpo
  - lugman-madhiai/Qwen3.5-9B-MHS-Interleaved
  - Hastagaras/Qwen3.5-9B-GLM-Wannabe
  - zenlm/zen4
  - trohrbaugh/Qwen3.5-9B-heretic-v2
tags:
  - merge
  - qwen3.5
  - terminal
  - cli
  - layer-wise-merge
  - gguf
  - quantized
  - llama-cpp
model_type: qwen3_5
language:
  - en
pipeline_tag: text-generation
library_name: transformers
---

# Qwen3.5-9B Terminal Merge

An experimental layer-wise merge of 16 Qwen3.5-9B variants aimed at terminal and CLI-style command generation.

## Status Update

I need to correct the original benchmark claim on this repo.

The first version of this model card reported results from an earlier evaluation setup that I no longer trust. After rerunning the comparison on the current 60-task Modal Harbor benchmark on March 17, 2026, this `v1` merge did not beat the current base `Qwen/Qwen3.5-9B` result.

I should not have presented the earlier result as a reliable improvement over base. Sorry about that.

For now, this repo should be treated as an experimental merge release rather than a confirmed upgrade over base Qwen.

## Current Internal Benchmark Snapshot

Fresh 60-task Modal Harbor run:

| Model | Tasks Passed | Pass Rate |
|-------|--------------|-----------|
| Qwen/Qwen3.5-9B | 53/60 | 88.3% |
| EganAI/qwen3.5-9b-terminal-merge | 50/60 | 83.3% |

The benchmark covers file operations, text processing, git workflows, networking, Python scripting, and system administration tasks in sandboxed execution environments.

## Download

| Format | Size | Link |
|--------|------|------|
| **Safetensors** (bf16) | ~18 GB | [Download](https://huggingface.co/EganAI/qwen3.5-9b-terminal-merge/tree/main) |
| **GGUF F16** | 17.9 GB | [Download](https://huggingface.co/EganAI/qwen3.5-9b-terminal-merge/blob/main/qwen3.5-9b-terminal-merge-f16.gguf) |
| **GGUF Q8_0** | 9.53 GB | [Download](https://huggingface.co/EganAI/qwen3.5-9b-terminal-merge/blob/main/qwen3.5-9b-terminal-merge-q8_0.gguf) |

## Model Details

- **Architecture**: Qwen3.5 (hybrid linear + full attention)
- **Parameters**: 9B total
- **Context Length**: 262,144 tokens
- **Precision**: bfloat16
- **Layers**: 32 (8 full attention + 24 linear attention)
- **Merge Method**: Layer-wise linear merge with optimized per-layer weights

## Source Models

This model combines optimized layer-wise weights from 16 Qwen3.5-9B variants spanning reasoning, instruction-following, and general capability specializations:

| Category | Models |
|----------|--------|
| **Core** | Qwen/Qwen3.5-9B, unsloth/Qwen3.5-9B |
| **Abliterated** | darkc0de/Qwen3.5-9B-heretic, lukey03/Qwen3.5-9B-abliterated, llmfan46/Qwen3.5-9B-ultimate-irrefusable-heretic, llmfan46/Qwen3.5-9B-ultra-heretic, jwest33/qwen3.5-9b-null-space-abliterated, trohrbaugh/Qwen3.5-9B-heretic-v2, osirisbrain/OsirisCortex-v6 |
| **Reasoning** | DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-THINKING, DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT, crownelius/Crow-9B-Opus-4.6-Distill-Heretic_Qwen3.5 |
| **Specialized** | alecccdd/Qwen3.5-9B-paraphrasing-orpo, lugman-madhiai/Qwen3.5-9B-MHS-Interleaved, Hastagaras/Qwen3.5-9B-GLM-Wannabe, zenlm/zen4 |

## How to Use

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "EganAI/qwen3.5-9b-terminal-merge"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="bfloat16",
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Find all Python files larger than 1MB and sort by size descending"}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### GGUF (llama.cpp)

```bash
# Download Q8_0 (recommended balance of quality/size)
wget https://huggingface.co/EganAI/qwen3.5-9b-terminal-merge/resolve/main/qwen3.5-9b-terminal-merge-q8_0.gguf

# Run with llama.cpp
./llama-cli -m qwen3.5-9b-terminal-merge-q8_0.gguf \
  -p "Find all files modified in the last 24 hours and sort by size:" \
  -n 512 --temp 0.7
```

### GGUF (Python)

```python
from huggingface_hub import hf_hub_download
from llama_cpp import Llama

model_path = hf_hub_download(
    repo_id="EganAI/qwen3.5-9b-terminal-merge",
    filename="qwen3.5-9b-terminal-merge-q8_0.gguf",
)

llm = Llama(model_path=model_path, n_ctx=4096, n_gpu_layers=-1)

output = llm(
    "List all running Docker containers and their memory usage:",
    max_tokens=256,
    temperature=0.7,
)
print(output["choices"][0]["text"])
```

### vLLM Serving

```bash
vllm serve EganAI/qwen3.5-9b-terminal-merge \
    --language-model-only \
    --dtype bfloat16 \
    --max-model-len 8192
```

> **Note**: Use `--language-model-only` since this is a multimodal architecture served for text-only inference.

## Training Details

The per-layer merge weights were optimized by evaluating candidates on a suite of 60 terminal tasks using vLLM inference in sandboxed environments. The optimization searched across layer-group weight distributions to find a strong blend of the 16 source models.

That earlier optimization run produced a candidate worth publishing as an experiment, but the original benchmark comparison against base should be considered superseded by the newer reevaluation above.

## Limitations

- This v1 release does not currently show a verified win over base `Qwen/Qwen3.5-9B` on my latest internal benchmark run
- Optimized specifically for terminal/CLI tasks; general-purpose performance may vary
- Requires `--language-model-only` flag when serving with vLLM due to multimodal architecture
- Visual capabilities are inherited from the base model but were not part of the optimization target