README.md

---
language:
- en
library_name: transformers
license: apache-2.0
license_link: https://huggingface.co/Qwen/Qwen3.6-35B-A3B/blob/main/LICENSE
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3.6-35B-A3B
tags:
- qwen3.6
- multimodal
- mixture-of-experts
- agentic-work
- occupational-work
- tool-use
---

# Fx-Work

Fx-Work is a 35B-class multimodal language model for professional work in realistic file-and-tool environments. It is designed to interpret heterogeneous work materials, reason over operational constraints, use tools when required, and produce deliverables that satisfy explicit professional requirements.

Fx-Work is built on Qwen3.6-35B-A3B and post-trained with 20K tool-interaction trajectories. The trajectories are generated from occupational work constructed with real-world artifacts, occupational knowledge, actionable requests, and itemized evaluation criteria. Execution-guided verification is used during data construction to identify and repair inconsistencies between the work request, reference materials, and rubric.

## Main Results

Fx-Work achieves the strongest results among the evaluated models at or below the 35B scale on all five reported metrics. It also surpasses DeepSeek-V4-Pro-Preview (1.6T) on four of the five metrics.

<p align="center">
  <a href="https://huggingface.co/endless-frontier/Fx-Work/blob/main/main_figure_fx_work.pdf">
    <img src="https://huggingface.co/endless-frontier/Fx-Work/resolve/main/main_figure_fx_work.png"
         alt="Fx-Work benchmark results" width="100%">
  </a>
</p>

## Benchmark Results

Results on GDPvalAA-v2, APEX-Agents-AA, and JobBench. GDPval Rubric is the normalized rubric score; GDPval Elo uses the August 4, 2026 snapshot; APEX-Agents reports pass@1; and JobBench reports the Main and Easy splits. AVG is the mean of normalized GDPval Elo, APEX-Agents-AA, and JobBench Main. `--` indicates that a result was not publicly available or was not tested. Bold marks the best value among the models at or below the 35B scale.

| Model | Parameters | GDPval Rubric | GDPval Elo | APEX-Agents pass@1 | JobBench Main | JobBench Easy | AVG |
|:--|--:|--:|--:|--:|--:|--:|--:|
| **Frontier models** |  |  |  |  |  |  |  |
| Qwen3.8-Max | 2.4T-A95B | 91.53 | 1739 | 38.9 | 52.57 | 82.81 | 51.14 |
| Kimi-K3 | 2.8T-A104B | 91.64 | 1687 | 35.5 | 43.81 | 80.72 | 46.22 |
| GPT-5.5 | -- | 90.50 | 1491 | 29.9 | 30.10 | 78.22 | 36.52 |
| GLM-5.2 | 753B-A40B | 90.34 | 1510 | 29.7 | 33.79 | 75.49 | 37.80 |
| DeepSeek-V4-Pro-Preview | 1.6T-A49B | 87.17 | 1304 | 19.2 | 18.90 | 64.08 | 26.10 |
| DeepSeek-V4-Flash-Preview | 284B-A13B | 86.49 | 1189 | 15.2 | 18.62 | 63.13 | 22.75 |
| **Models at or below 35B** |  |  |  |  |  |  |  |
| Nex-N2-mini | 35B-A3B | 69.96 | 1065 | 15.6 | 9.39 | 52.93 | 17.74 |
| Agents-A1 | 35B-A3B | 74.31 | 877 | 11.7 | 7.19 | 41.74 | 12.58 |
| Apodex-1.0-mini | 35B-A3B | 78.36 | 973 | 14.6 | 7.97 | 44.38 | 15.40 |
| Occamy-1.0 | 35B-A3B | 79.50 | 1200 | 20.5 | 18.88 | 65.49 | 24.79 |
| Qwen3.6-27B | 27B | 85.10 | 1138 | 16.6 | 19.18 | 64.59 | 22.56 |
| Qwen3.6-35B-A3B | 35B-A3B | 82.61 | 1053 | 13.4 | 14.99 | 56.02 | 18.68 |
| **Fx-Work** | **35B-A3B** | **86.66** | **1352** | **25.2** | **25.19** | **70.06** | **31.00** |



## Quickstart

For streamlined integration, we recommend serving Fx-Work through an OpenAI-compatible API using SGLang or vLLM. The repository requires Hugging Face authentication when access is restricted.

> [!Important]
> Fx-Work uses a native context length of 262,144 tokens. If GPU memory is limited, reduce the context length; for long-horizon professional work, retaining at least 128K tokens is recommended.

### SGLang

[SGLang](https://github.com/sgl-project/sglang) provides an OpenAI-compatible server for high-throughput inference.

```bash
python -m sglang.launch_server \
  --model-path endless-frontier/Fx-Work \
  --port 8000 \
  --tp-size 8 \
  --mem-fraction-static 0.8 \
  --context-length 262144 \
  --reasoning-parser qwen3
```

### vLLM

[vLLM](https://github.com/vllm-project/vllm) provides a high-throughput OpenAI-compatible server.

```bash
vllm serve endless-frontier/Fx-Work \
  --port 8000 \
  --tensor-parallel-size 8 \
  --max-model-len 262144 \
  --reasoning-parser qwen3
```

Both commands expose an endpoint at `http://localhost:8000/v1`. Tool execution is provided by the surrounding agent harness; the model server exposes generation and, where configured, tool-call responses.

### Transformers

```python
import torch
from transformers import AutoProcessor, Qwen3_5MoeForConditionalGeneration

model_id = "endless-frontier/Fx-Work"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3_5MoeForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {
        "role": "user",
        "content": "Summarize the key risks and next actions in this work request.",
    }
]
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
).to(model.device)

with torch.inference_mode():
    output_ids = model.generate(**inputs, max_new_tokens=1024)

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

The processor supports the multimodal image and video message format provided by Qwen3.6.


## Citation

If you use Fx-Work, please cite:

```bibtex
@article{zhu2026workgenesis,
  title={WorkGenesis: Building the Worlds That Teach Agents to Work},
  author={Zhu, Xinyu and Liu, Fenyi and Cai, Yuzhu and Tang, Shuo and Ye, Rui and Zhang, Linfeng and Chen, Siheng},
  journal={arXiv preprint arXiv:2609.39325},
  year={2026}
}
```

## License

Fx-Work is released under the Apache 2.0 license, subject to the terms of the base model and third-party components. See the [Qwen3.6-35B-A3B license](https://huggingface.co/Qwen/Qwen3.6-35B-A3B/blob/main/LICENSE) for the upstream terms.