Introducing Hy4 Preview
- 原文链接: https://simonwillison.net/2026/Aug/29/hy4/
- 作者: Simon Willison
- 日期: 2026-08-29
- 分类: models
- 来源类型: article
- 标签: open-weights, tencent, moe, long-context, reasoning-models
- 质量评分: 3/5
- 抓取时间: 2026-08-30T23:41:49+08:00
中文导读
腾讯发布开源权重 Hy4 Preview：770B 总参 / 49B 激活，1M token 上下文，Hugging Face 权重 1.56TB，纯文本输入无视觉；相比 7 月的 Hy3（295B 总参 / 21B 激活 / 256K 上下文 / 598GB）是全面放大。作者读 chat_template.jinja 确认推理控制只有两档：reasoning_effort 取 high（默认）或 no_think，即开推理或关推理的二值设计。用 OpenRouter 跑 pelican-SVG 测试，推理轨迹呈明显截断的英语（Maybe add water? no.），作者判断是隐藏推理文本的 token 效率取舍，与 Goedecke 描述的 Claudish 现象同源。对跟踪国产开源 MoE 规模曲线与推理控制接口设计的人，这是难得的一次性规格核对。
原文（抓取存档 · AK-RSS source archive (crawled 2026-08-30)）
# Introducing Hy4 Preview
> 作者: Simon Willison
> 原文链接: https://simonwillison.net/2026/Aug/29/hy4/
---
29th August 2026 - Link Blog
**[Introducing Hy4 Preview](https://hy.tencent.ai/research/hy4-preview)**. New open weight text input (no vision) LLM from Chinese company Tencent today: 770B total parameters, 49B active parameters, 1M token context window, [1.56TB on Hugging Face](https://huggingface.co/tencent/Hy4-preview).
This is a big size increase from their previous [Hy3](https://huggingface.co/tencent/Hy3) in July, which was 295B, 21B active, 256,000 context, 598GB.
I recently started using model chat templates to better understand their capabilities. Here's Hy4's [chat\_template.jinja](https://huggingface.co/tencent/Hy4-preview/blob/main/chat_template.jinja) on Hugging Face, which includes this section:
{%- 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 it looks like there are just two reasoning effort levels: "high" (the default) and "no\_think" (reason by disabled).
I tried my "Generate an SVG of a pelican riding a bicycle" prompt with the default high reasoning [via OpenRouter](https://openrouter.ai/tencent/hy4-preview#apps) and [got this](https://tools.simonwillison.net/markdown-svg-renderer#url=https%3A%2F%2Fgist.github.com%2Fsimonw%2Fcb69816b3fb940f2782569a82a523af1):
Quoting the reasoning trace:
> \[...\] Let's maybe add a helmet? It could improve riding theme, but may obscure head. Maybe a small cycling cap or helmet? The user didn't ask; can add red helmet? Might be cute. But pelican with big beak; a helmet might obscure. Better maybe no.
>
> Maybe add sunglasses? no.
>
> Maybe add water? no.
It's interesting how the reasoning trace uses slightly truncated English, presumably because perfect grammar isn't useful or token efficient for hidden reasoning text.
Posted [29th August 2026](https://simonwillison.net/2026/Aug/29/) at 11:53 pm