Fx-Work-1028

Fx-Work-1028 is an improved version of Fx-Work, built to handle everyday professional work with stronger reasoning, tool use, and task completion. It is a 35B-class multimodal mixture-of-experts model built on Qwen3.6-35B-A3B, with approximately 3B activated parameters.

We further strengthen Fx-Work's ability to carry out practical work in realistic file-and-tool environments. Fx-Work-1028 is designed to understand work requests, interpret documents and other reference materials, reason over operational constraints, and use tools to produce useful deliverables. Its intended tasks include preparing documents and reports, analyzing spreadsheets and business data, organizing information, and completing workflows that involve multiple files and tools.

With only 35B parameters, Fx-Work-1028 delivers work-task performance comparable to GLM-5.2 across the five benchmarks reported below. It improves over the original Fx-Work on every benchmark, with particularly large gains on AutomationBench and Claw-Eval.

Main Results

Fx-Work-1028 scores higher than GLM-5.2 on JobBench Main, AutomationBench, and Claw-Eval, while remaining close on GDPvalAA-v2.1 Elo (1328 vs. 1374) and APEX-Agents-AA (28.2 vs. 29.7):

Benchmark Results

Higher is better for every metric. GDPvalAA-v2.1 reports Elo; APEX-Agents-AA and AutomationBench report pass@1; JobBench reports its Main score; and Claw-Eval reports pass³. Pass metrics are shown as percentages.

Model	GDPvalAA-v2.1 Elo	APEX-Agents-AA pass@1	JobBench Main	AutomationBench pass@1	Claw-Eval pass³
Frontier models
GLM-5.3	1651	35.2	48.48	48.20	72.2
Qwen3.8-max	1608	38.9	52.57	51.30	72.8
DeepSeek V4.1 Flash-0910	1600	38.4	49.41	54.80	64.3
Kimi K3	1533	35.5	43.81	46.70	69.3
GLM-5.2	1374	29.7	33.79	20.30	67.9
Models at or below 35B
Sage-A1.5	675	14.8	14.32	8.78	65.0
Apodex-1.1-mini	883	23.5	16.31	14.30	45.1
Qwen3.6-35B-A3B	895	13.4	14.99	9.30	49.7
Occamy-1.0	1088	20.5	18.88	27.60	71.1
Nex-N2.5-mini	1304	17.0	28.50	32.30	36.1
Fx-Work	1248	25.2	25.19	18.00	38.9
Fx-Work-1028	1328	28.2	34.34	43.33	70.1

Quickstart

For integration with an agent workflow, we recommend serving Fx-Work-1028 through an OpenAI-compatible API using SGLang or vLLM. The examples use the model repository endless-frontier/Fx-Work-1028. Hugging Face authentication is required if repository access is restricted.

Fx-Work-1028 follows Fx-Work's 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 provides an OpenAI-compatible server for high-throughput inference.

python -m sglang.launch_server \
  --model-path endless-frontier/Fx-Work-1028 \
  --port 8000 \
  --tp-size 8 \
  --mem-fraction-static 0.8 \
  --context-length 262144 \
  --reasoning-parser qwen3

vLLM

vLLM provides a high-throughput OpenAI-compatible server.

vllm serve endless-frontier/Fx-Work-1028 \
  --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

import torch
from transformers import AutoProcessor, Qwen3_5MoeForConditionalGeneration

model_id = "endless-frontier/Fx-Work-1028"
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-1028, please cite:

@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-1028 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 for the upstream terms.