Model reference · open weights

WebDancer

LLMs Alibaba-NLP Text gen 1 build Open weights 3k dl/mo

WebDancer is an open-weight language model from Alibaba. WebDancer-32B (BF16) weighs 65.5 GB; the smallest configuration that runs it is H100 80 GB.

WebDancer is a 32.8B parameter text-generation model developed by Alibaba for autonomous information seeking and agentic search reasoning.
It utilizes a ReAct framework and was trained using a four-stage paradigm that includes supervised fine-tuning and reinforcement learning.
The model supports a context length of 40,960 tokens and is released under the MIT licence.

Summary of the Alibaba-NLP/WebDancer-32B model card, 2026-10-04

What it is

What it runs on

Memory and cards for WebDancer-32B (BF16)

65.5 GBweights, file size

262 MBcache per 1K tokens

896 MBruntime overhead, at least

40,960 tokenscontext max

Card	Requests at once		Context max	Memory
Card	8K each	32K each	Context max	Memory
RTX 3060 12 GB … L40S 48 GB 6 smaller cards	—	—	—
A100 80 GB	5	1	all 40K	78.2 GB
H100 80 GB	3	—	25K	78.1 GB
RTX PRO 6000 Blackwell 96 GB	10	2	all 40K	93.8 GB
DGX Spark (GB10) 128 GB unified	16	4	all 40K	107 GB
H200 141 GB	31	7	all 40K	138 GB
B200 180 GB	48	12	all 40K	176 GB
2× L40S 48 GB tensor parallel	9	2	all 40K	44.0 GB a card
4× RTX 4090 24 GB tensor parallel	11	2	all 40K	23.4 GB a card
4× RTX 3090 24 GB tensor parallel	11	2	all 40K	23.4 GB a card
4× RTX 5090 32 GB tensor parallel	25	6	all 40K	31.0 GB a card
2× H100 80 GB tensor parallel	36	9	all 40K	78.1 GB a card
2× A100 80 GB tensor parallel	41	10	all 40K	78.2 GB a card

Memory needed at each load

One card, with vLLM's small-card settings.

From the model card

Read the model card

You can download the model then run the inference scipts in https://github.com/Alibaba-NLP/WebAgent.

Native agentic search reasoning model using ReAct framework towards autonomous information seeking agency and Deep Research-like model.
We introduce a four-stage training paradigm comprising browsing data construction, trajectory sampling, supervised fine-tuning for effective cold start, and reinforcement learning for improved generalization, enabling the agent to autonomously acquire autonomous search and reasoning skills.
Our data-centric approach integrates trajectory-level supervision fine-tuning and reinforcement learning (DAPO) to develop a scalable pipeline for training agentic systems via SFT or RL.
WebDancer achieves a Pass@3 score of 61.1% on GAIA and 54.6% on WebWalkerQA.

Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.

Details