---
license: apache-2.0
base_model: Qwen/Qwen3.5-9B
library_name: transformers
pipeline_tag: image-text-to-text
tags:
  - softwaregym
  - computer-use
  - gui-agent
  - multimodal
  - fine-tuned
---

# SoftwareGym — Qwen3.5-9B, Step 1545

This repository contains the full SoftwareGym fine-tuned checkpoint at training step **1545**, based on **Qwen/Qwen3.5-9B**.

## Model size

| Property | Value |
| --- | --- |
| Architecture | `Qwen3_5ForConditionalGeneration` |
| Parameters | 9,409,813,744 (approximately 9.41B) |
| Weight precision | BF16 (`bfloat16`) |
| Weight files | 4 Safetensors shards and their index |
| Total weight size | 18,819,721,120 bytes (approximately 18.82 GB / 17.53 GiB) |
| Checkpoint | Step 1545 |

Tokenizer, processor, generation configuration, and the chat template are included with the weights.

## Deployment parameters

The checkpoint's reference deployment uses **MS-Swift with the Transformers inference backend**. Its recorded software environment is:

- Python 3.10.19
- PyTorch 2.8.0 with CUDA 12.8
- Transformers 5.10.2
- MS-Swift 4.4.0.dev0
- FlashAttention 2.8.3
- Accelerate 1.10.1

These are the versions recorded for the original serving environment, rather than a separately tested minimum-version requirement.

| Parameter | Reference value |
| --- | --- |
| Inference backend | `transformers` |
| Data type | `bfloat16` |
| Attention implementation | `flash_attn` |
| Device map | `auto` |
| Maximum sequence length | 65,536 tokens |
| Maximum new tokens | 8,192 |
| Maximum batch size per replica | 1 |
| Sampling temperature in the reference request | 1.0 |
| Top-p in the reference request | 0.95 |

After downloading this repository to a local directory, start one replica with the reference serving settings:

```bash
CUDA_VISIBLE_DEVICES=0 swift deploy \
  --model /path/to/SoftwareGym \
  --infer_backend transformers \
  --served_model_name softwaregym-step1545 \
  --host 127.0.0.1 --port 8450 \
  --torch_dtype bfloat16 \
  --attn_impl flash_attn \
  --device_map auto \
  --max_length 65536 \
  --max_new_tokens 8192 \
  --max_batch_size 1 \
  --no_verbose
```

The original deployment used **8 NVIDIA A800-SXM4-80GB GPUs**, with **one independent replica per GPU**, backend ports **8450–8457**, and an OpenAI-compatible round-robin gateway on port **8441**. This is an eight-replica deployment, rather than eight-way tensor parallelism. The command above starts one replica; the gateway is external to the model files in this repository.

For GUI-agent use, provide the task, screenshot, and interaction history through the included chat template. The reference response format contains `<think>`, `## Action:`, and `## Code:` sections. Pointer coordinates are integers in the **0–999** range and must be mapped to screen pixels by the GUI executor.

## License

This repository uses Apache-2.0, consistent with the repository's existing license metadata and the Qwen3.5-9B base model. See the included `LICENSE` for the license text.