twenty ninth August 2026 – Hyperlink Weblog
Introducing Hy4 Preview. New open weight textual content enter (no imaginative and prescient) LLM from Chinese language firm Tencent in the present day: 770B complete parameters, 49B energetic parameters, 1M token context window, 1.56TB on Hugging Face.
It is a large measurement enhance from their earlier Hy3 in July, which was 295B, 21B energetic, 256,000 context, 598GB.
I just lately began utilizing mannequin chat templates to higher perceive their capabilities. Here is Hy4’s chat_template.jinja on Hugging Face, which incorporates this part:
{%– set reasoning_effort = ‘excessive’ %}
{%– elif reasoning_effort not in [‘high’, ‘no_think’] %}
{%– if reasoning_effort is none %}
{{- raise_exception(‘reasoning_effort error : None, needs to be no_think/excessive’) }}
{%– else %}
{{- raise_exception(‘reasoning_effort error : ‘ + reasoning_effort + ‘, needs to be no_think/excessive’) }}
{%– endif %}
{%– endif %}
So it seems to be like there are simply two reasoning effort ranges: “excessive” (the default) and “no_think” (purpose by disabled).
I attempted my “Generate an SVG of a pelican driving a bicycle” immediate with the default excessive reasoning through OpenRouter and acquired this:
Quoting the reasoning hint:
[…] Let’s possibly add a helmet? It might enhance driving theme, however could obscure head. Possibly a small biking cap or helmet? The person did not ask; can add purple helmet? Could be cute. However pelican with large beak; a helmet would possibly obscure. Higher possibly no.
Possibly add sun shades? no.
Possibly add water? no.
It is attention-grabbing how the reasoning hint makes use of barely truncated English, presumably as a result of excellent grammar is not helpful or token environment friendly for hidden reasoning textual content.
