Skip to content
4All API Docs
Main Site News Console

Hy4 Preview Is Now Available

· Simon Willison Translated
技巧LLM

Hy4 Preview Officially Released

Tencent, a Chinese company, today released a new open-weight LLM that accepts text-only input (vision is not supported): it has 770 billion total parameters, 49 billion active parameters, and supports a 1-million-token context window. The model occupies 1.56 TB on Hugging Face.

Compared with Hy3, which was released in July, this model is substantially larger. Hy3 has 295 billion total parameters, 21 billion active parameters, a 256,000-token context window, and a model size of 598 GB.

I recently started studying models’ chat templates to better understand their capabilities. Here is Hy4’s chat_template.jinja on Hugging Face, which includes the following:

{%- if not reasoning_effort is defined %}
    {%- set reasoning_effort = 'high' %}
{%- elif reasoning_effort not in ['high', 'no_think'] %}
    {%- if reasoning_effort is none %}
        {{- raise_exception('reasoning_effort error : None, should be no_think/high') }}
    {%- else %}
        {{- raise_exception('reasoning_effort error : ' + reasoning_effort + ', should be no_think/high') }}
    {%- endif %}
{%- endif %}

So this appears to mean that it has only two reasoning-effort levels: “high” (the default) and “no_think” (reasoning disabled).

Using the default high reasoning effort, I tried the prompt “Generate an SVG of a pelican riding a bicycle” through OpenRouter, and got the following result:

A flat vector cartoon illustration of a white pelican with a huge orange beak riding a red bicycle to the right along a gray road with a white dashed center line; its orange webbed feet are on the pedals, and its gray tail feathers fan out behind it. The background is a pale blue sky with a yellow sun, white clouds, and horizontal white motion lines indicating speed.

Here is an excerpt from its reasoning process:

[…] Perhaps add a helmet? This would reinforce the cycling theme, but might cover the head. Maybe a small cycling cap or helmet? The user didn’t ask for one; adding a red helmet could be cute. But the pelican has a large beak, and the helmet might obstruct it. Better not to add one.

Perhaps add sunglasses? No.

Perhaps add some water? No.

Interestingly, this reasoning process uses somewhat incomplete English phrasing. Presumably, this is because perfect grammar has no practical value for hidden reasoning text and would instead waste tokens.