Tongyi DeepResearch

[spare island]

Pls https://huggingface.co/Alibaba-NLP/Tongyi-DeepResearch-30B-A3B

🙏

[brave shale]

I second this motion

[fossil tundra]

this is live! https://openrouter.ai/alibaba/tongyi-deepresearch-30b-a3b

[viscid swan]

alibaba puts a heavy premium on their endpoints

[solemn spruce]

lovely pricing

[proud edge]

oh this would be nice to have, would like to try deep research models with open deep research platforms

[spare island]

Let’s gooo

Super cool

[bitter sentinel]

i don’t understand this

is this supposed to search the web? how do we make it do that

[viscid swan]

Give your own tools

It's good at using it

[limpid fox]

This fails at tool calling for me for some reason xD

[limber cedar]

Yeah @fossil tundra Is this natively supposed to have access to web search or do we need to provide it a web search tool?

[fossil tundra]

It doesn’t have native web search, but you can add :online to to the end of the model name to enable Exa-powered web search

https://openrouter.ai/docs/features/web-search

[limber cedar]

oh my god this model is incredible now

it was hallucinating every url before and i didn't understand why but that makes total sense now

[fossil tundra]

Nice :)

[limber cedar]

Def getting some errors on some requests though like

(Tongyi request failed)

or

We need to perform a new web search. Let's use the appropriate tool.
</think>

We to=assistant<|channel|>commentary<|message|>Error: Tool call is not valid.
</tool_call>
</tool_response

[chrome stratus]

This model is not for general purpose

It's an agentic deep research model that makes use of tools and is really good at assimilating sources and making queries for deep research

It needs to be paired with a deep research platform/code

[south fern]

I’d suggest checking out the Alibaba Cloud Developer Community on Discord,
the admin Lara usually loops in the Qwen team to help solve issues directly.
It’s been pretty helpful for a lot of people

[limber cedar]

Are there any examples of this? Curious what kind of harness is required for this. Is it just a while loop running over and over again checking the queries its generated and the tool calls its made until a certain condition is met?

[limber cedar]

Ah, will do!

[chrome stratus]

The best way to learn about the procedure is through their own report:
https://github.com/Alibaba-NLP/DeepResearch/

This github repo has the pipeline required in here, they also mention how you can connect it via openrouter api

So, inorder to use this model effectively, you need to download this repo and follow the steps mentioned in this repo and then use the model with this pipeline(It already has the code and everything, you probably just need to plug in some api key and do some minor changes)