DataWorld-35B-A3B

DataWorld-35B-A3B

A Qwen3.6-35B-A3B model tuned with multi-turn agentic reinforcement learning on the DataWorld tasks, for tool-calling agents. For teams serving a large mixture-of-experts model.

Made by Zhanli-Li on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: DataWorld-35B-A3B is a reinforcement-learning fine-tune of Qwen3.6-35B-A3B, a mixture-of-experts model with 35B parameters in total and 3B active per token. It was trained on tool-calling agent runs from the DataWorld task suite. The repository holds the resulting checkpoint in the standard Hugging Face safetensors format.

Who it suits: It suits teams that serve a large open model for agents that call tools and work on software or data tasks.

Industry | Who uses it | How they use it | Example
Software | Agent developers | Serve it as the model behind a tool-calling agent | An agent that edits files and runs commands
Data work | Data teams | Use an agent tuned on data environment tasks | An agent that works through a folder of data files
AI infrastructure | Serving engineers | Deploy it with the same recipe as the base model | Launching it with the base model's server options
Research | Reinforcement learning researchers | Study an agentic RL fine-tune against its base | Comparing it with the original Qwen3.6 model

You can skip it if you have a single small GPU, since the weights are about 70 GB and the card shows a server command that uses eight GPUs.

Where to use it:

• SGLang or vLLM for serving, which the card names with minimum versions
• The transformers library, in a version that knows the qwen3_5_moe architecture
• Multi-GPU servers, since the card's server example splits the model across eight devices
• Windows, macOS or Linux, with Python for the download

Use case | Who | What they get back
Tool-calling agents | Developers | Tool calls and replies from the model
Image and text input | Application teams | Multimodal answers, since the vision tower is included
Long-context work | Researchers | A window of up to 262,144 tokens, or more with the long-context overrides

How to set it up:

• sglang 0.5.10 or newer, vllm 0.19.0 or newer, or a transformers version that knows the architecture
• A GPU server, since the BF16 weights are about 70.2 GB
• Python and pip for the download

1. Install the Hugging Face hub package with pip.
2. Download the repository files, or point your server at the model path.
3. Start a server with the reasoning and tool-call parsers the card names, or load it with transformers.

pip install -U huggingface_hub
hf download Zhanli-Li/DataWorld-35B-A3B

How to use it: You serve the model the way the base Qwen3.6 model is served, changing only the model path. The card says to keep the reasoning and tool-call parsers on, because the model was trained on tool-calling runs.

python -m sglang.launch_server \
  --model-path Zhanli-Li/DataWorld-35B-A3B \
  --port 8000 --tp-size 8 \
  --mem-fraction-static 0.8 --context-length 262144 \
  --reasoning-parser qwen3 --tool-call-parser qwen3_coder

To load it directly in Python with transformers, the card gives this.

from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "Zhanli-Li/DataWorld-35B-A3B"   # or a local path
model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(repo)

Example: A team launches the server with the command above and connects an agent that has file and shell tools. The agent sends a task, the model replies with a tool call, and the agent runs it and returns the result. The card says the model was trained on exactly this kind of loop.

For newcomers and experts: If you are new to this, treat it as a server-grade model. Read the base model's serving guide first, and keep the parser options from the card turned on.

If you know the field, the architecture is the base model's: 40 layers, 256 experts with top-8 routing, hybrid linear and full attention, and a 262,144-token context. The vision tower is included. The multi-token prediction block is not, so the base model's speculative decoding recipes do not apply.

Download the checkpoint
Start a server with the parsers on
Connect an agent with tools
The model replies with tool calls
The agent runs them and returns results

Good to know:

Item | Detail
Total and active parameters | 35B total, 3B active per token
Size on disk | About 70.2 GB in 14 shards
Vision | The vision tower is included
Speculative decoding | Not available, since the prediction block is absent

• Limits: the card shows no benchmark numbers, and it needs a multi-GPU server for its sample setup.
• Parsers: keep the reasoning and tool-call parsers enabled, as the card prescribes.
• License: Apache 2.0, the same as the base model.
• Cost to run: none named.
• Data it reaches: it copies large model files to your computer, and any tool access depends on the agent you connect.

On Hugging Face, 14 November 2026

• Task: reinforcement learning
• Library: transformers
• Parameters: 35.1B
• Weights: about 70.2 GB in BF16; about 17.6 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 140
• Likes: 20
• License: apache-2.0
• Based on: Qwen/Qwen3.6-35B-A3B
• Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, agent, reinforcement-learning, tool-use

Links

• Model: https://huggingface.co/Zhanli-Li/DataWorld-35B-A3B
• Files: https://huggingface.co/Zhanli-Li/DataWorld-35B-A3B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Zhanli-Li/DataWorld-35B-A3B

Then follow the usage in its model card. Nothing in this kit runs by itself.