Change default InferenceClient model to Qwen/Qwen3-Next-80B-A3B-Thinking

[suryabdev]

Follow up of #1801 and a Fix for #1808,

The default model for InferenceClientModel is Qwen/Qwen2.5-Coder-32B-Instruct. It does not work because the current default provider doesn't support tool calling (More details in the issue)
This PR changes the default model to Qwen/Qwen3-Next-80B-A3B-Thinking

[suryabdev]

cc: @albertvillanova / @aymeric-roucher please review when you are free

[aymeric-roucher]

@suryabdev Qwen3-Coder isn't that good in my experience : Qwen/Qwen3-Next-80B-A3B-Thinking worked better (but I only tried a few runs)

[suryabdev]

@aymeric-roucher I've only run some basic tests myself to test functionality. Haven't run any benchmarks like GAIA
I've changed the model to Qwen/Qwen3-Next-80B-A3B-Thinking

[suryabdev]

Just something to note, the price of the default model will increase, but all providers support tool calling (Doc). From

to

[albertvillanova]

To chose a new default model, I would recommend benchmarking the candidates and selecting based on their performance.

Another possibility would be to keep the current default model and just add the associated provider that was working until now.

[suryabdev]

I would recommend benchmarking the candidates and selecting based on their performance

@albertvillanova That is fair, I haven't run the benchmarks before. Let me try to run them now
https://github.com/huggingface/smolagents/blob/main/examples/smolagents_benchmark/run.py

Another possibility would be to keep the current default model and just add the associated provider that was working until now.

I don't think we should change the default provider for the InferenceClientModel, That might impact situations when a user tries to a different model and doesn't set the provider.

smolagents/src/smolagents/models.py

Line 1416 in 8f4dc91

provider: str | None = None,

The InferenceClient has good auto-picking behavior to choose the cheapest provider.
Adding conditonal logic that adds the provider only if the model is Qwen/Qwen2.5-Coder-32B-Instruct or None could work but doesn't feel very clean to me

[aymeric-roucher]

I would be strongly in favor of updating the default model in InferenceClient, as everywhere else : the seed is random anyway so there shouldn't be any conditional logic based on using Qwen/Qwen2.5-Coder-32B-Instruct rather than another model.
Plus Qwen3 series of model is just a net improvement over 2.5 in all aspects : latency, performance, price. So let's not lock ourselves into another model. And anyway, providers will end up discontinuing 2.5 probably before 2026, as new models keep rolling in.

[suryabdev]

Ran the benchmark only for the CodeAgent. Qwen/Qwen3-Next-80B-A3B-Thinking is an improvement over Qwen/Qwen2.5-Coder-32B-Instruct

Found some bugs while running the benchmark script so I raised a PR #1822.
I had some questions on running the benchmark for the ToolCallingAgent which I mentioned in that PR

[suryabdev]

the seed is random anyway so there shouldn't be any conditional logic based on using Qwen/Qwen2.5-Coder-32B-Instruct rather than another model.

@aymeric-roucher sorry could you elaborate. I didn't fully understand. Do you mean the InferenceClient randomly picks a provider?

[aymeric-roucher]

@suryabdev I meant that changing the model should not break working pipelines for users (except if their pipeline has an assert check on the model_id), because there's no expectation of reproducibility anyway when using generation.
So tests likeassert that my pipelines outputs exactly "Hi, I'm an assistant and the answer is A" won't exist (they would be broken by randomness), so we don't have to fear that updating the model could break working pipelines.

[aymeric-roucher]

Thank you @suryabdev ! 😃 Only need to fix conflicts before going ahead!

[suryabdev]

@aymeric-roucher Thanks for the review! I resolved the merge conflicts
You can trigger the PR checks when you are free

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[albertvillanova]

Thanks! We should update the PR title and description accordingly.

[albertvillanova]

Oh, it seems the default model was already changed before this PR! 😲

[suryabdev]

It was changed in this PR #1801 yesterday, But since there was no release yet no users will be impacted

[suryabdev]

We should update the PR title and description accordingly.

@albertvillanova thanks for the review, I've updated the PR title and description
Please merge when you are free